A function call-based gas turbine fault quantification diagnosis method and system
By constructing an intelligent agent architecture based on a large language model and a service-oriented operation and maintenance algorithm cluster, the problems of lack of engineering constraints in the model and fixed algorithm process in gas turbine fault diagnosis are solved. This enables automated, modular, and interpretable diagnosis of gas turbine faults, improving diagnostic accuracy and adaptability to complex faults.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNITED GAS TURBINE TECH CO LTD
- Filing Date
- 2026-03-31
- Publication Date
- 2026-06-05
AI Technical Summary
Existing gas turbine fault diagnosis technologies often lack general-purpose large language models that are constrained by engineering mechanisms, which can easily lead to uninterpretable reasoning results. Traditional fault diagnosis algorithms have limited functionality and fixed processes, making it difficult to dynamically combine and call them according to different abnormal scenarios. Furthermore, natural language fault descriptions and professional operation and maintenance algorithms lack automatic connection and process-level scheduling capabilities.
We construct an intelligent agent architecture with a large language model as the core scheduling engine and a service-oriented operation and maintenance algorithm cluster as the computing core. By building a structured knowledge base and an intelligent agent collaborative scheduling mechanism, we can achieve structured parsing of natural language input, automatic planning of algorithm flow, multi-module collaborative computing, and interpretable output of results. Combining knowledge graphs and physical mechanisms in the field of gas turbine operation and maintenance, we can perform automated scheduling of function calls and algorithm modules.
It improves the accuracy, quantification, and process flexibility of gas turbine fault diagnosis, enhances the adaptability to complex coupled fault scenarios, realizes the automation, modularization, and interpretability of the diagnosis process, and improves the engineering reliability and interpretability of the diagnosis results.
Smart Images
Figure CN122153081A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to fault diagnosis technology for gas turbines, specifically to a method and system for quantitative fault diagnosis of gas turbines based on function calls. Background Technology
[0002] As a core power source in modern power systems, the operating status of gas turbines directly affects energy conversion efficiency, power supply reliability, and equipment operational safety. Gas turbines typically operate under complex conditions of high temperature, high pressure, and high speed for extended periods. Internal components, such as turbine blades, bear significant thermal loads and mechanical stresses. Performance degradation, component wear, or control anomalies can easily lead to efficiency decline, excessive emissions, or even unplanned shutdowns. In large generator sets, gas turbine failures not only result in equipment damage but can also impact grid stability and economic operation. Therefore, real-time monitoring and accurate fault diagnosis of gas turbine operating status, timely identification of potential anomalies, and provision of quantitative assessments and remedial recommendations are crucial technical aspects for ensuring equipment safety, extending service life, and reducing operation and maintenance costs.
[0003] Existing gas turbine fault diagnosis technologies mainly include methods based on physical mechanism models, methods based on rules and expert knowledge, and data-driven methods. Methods based on physical mechanisms typically establish thermodynamic models, component performance models, or state equation models to calculate and compare key operating parameters. While possessing a clear theoretical basis, these methods are complex to build and require high-quality parameter calibration and operating data, making it difficult to fully cover all operating scenarios under complex real-world conditions. Methods based on rules or expert knowledge rely on the long-term experience accumulated by maintenance personnel, using rule bases to match and judge abnormal phenomena. These methods are effective in well-structured typical fault scenarios, but they are highly dependent on experts, have high rule maintenance costs, and are difficult to adapt to complex coupled faults or novel anomaly patterns. In recent years, data-driven methods based on machine learning have been increasingly applied to gas turbine condition monitoring. These methods train historical data using neural networks, support vector machines, or ensemble learning models to achieve fault classification and trend prediction. However, these methods are highly dependent on the number of samples, have insufficient generalization ability when fault samples are scarce or operating conditions change significantly, and their internal decision-making mechanisms often lack interpretability, making it difficult to meet the reliability and traceability requirements of engineering applications. With the development of artificial intelligence technology, Large Language Models (LLMs) have demonstrated significant advantages in natural language understanding, knowledge integration, and complex task reasoning. Through pre-training on massive amounts of data, LLMs possess strong semantic parsing and cross-domain knowledge association capabilities, achieving good results in scenarios such as intelligent question answering, knowledge retrieval, and decision support. However, general-purpose LLMs typically lack engineering constraints and physical mechanism knowledge specific to the gas turbine field. Their outputs are mainly generated based on statistical language patterns, which carries the risk of "illusionary" reasoning and makes it difficult to guarantee the accuracy of numerical calculations and the engineering reliability of results. Furthermore, LLMs themselves lack the ability to perform complex numerical calculations or call professional algorithm tools, lacking reasonable function call mechanisms and process scheduling mechanisms, thus significantly limiting their application in industrial fault quantitative diagnosis. Therefore, how to retain the natural language understanding advantages of LLMs while introducing engineering algorithms and physical mechanism constraints to achieve automated scheduling and collaborative computation of professional algorithms has become an urgent problem to be solved in the application of LLMs in the industrial field.
[0004] Therefore, a technical solution that deeply integrates the task understanding and process planning capabilities of large language models with engineering-verified gas turbine operation and maintenance algorithms, and achieves automatic algorithm combination and quantitative calculation through a function call mechanism, can effectively solve the aforementioned gas turbine system fault diagnosis problem. By constructing a structured knowledge base, a service-oriented algorithm cluster, and an intelligent agent collaborative scheduling mechanism, the system possesses both the ability to understand natural language anomaly descriptions and the ability to perform accurate calculations and comprehensive analysis under the constraints of physical mechanisms and quantitative models, thereby achieving automation, modularization, and interpretability of the diagnostic process. This integrated technical approach not only improves the accuracy and stability of fault identification but also enhances the system's adaptability to complex coupled fault scenarios, providing more reliable and efficient technical support for intelligent operation and maintenance of gas turbines. Summary of the Invention
[0005] The technical problem addressed by this invention is the lack of engineering mechanism constraints in the general large language model, which easily produces uninterpretable reasoning results, and the limited functionality, fixed process, and difficulty in dynamically combining and calling different abnormal scenarios in existing gas turbine fault diagnosis technologies. Furthermore, it addresses the shortcomings of existing systems in automatically connecting natural language fault descriptions with professional operation and maintenance algorithms, lacking process-level scheduling capabilities and continuous optimization capabilities. This invention proposes a function call-based quantitative diagnosis method and system for gas turbine faults. By constructing an intelligent agent architecture with a large language model as the core scheduling engine and a service-oriented operation and maintenance algorithm cluster as the computing core, it achieves structured parsing of natural language input, automatic algorithm process planning, multi-module collaborative computation, and interpretable result output, thereby improving the accuracy, quantification, process flexibility, and engineering practicality of gas turbine fault diagnosis.
[0006] This invention is achieved through the following technical solution: The Large Language Model (LLM) in the system undertakes key tasks such as natural language parsing, task understanding, function call decision-making, and process planning. By fine-tuning the general pre-trained model with professional corpus from the gas turbine operation and maintenance field, the model's ability to understand terminology and business scenarios in this field is improved; simultaneously, knowledge augmentation processing is performed based on a knowledge graph from the gas turbine operation and maintenance field. The language model adapted by the above strategies can accurately parse user-input fault descriptions, extract fault information and associated tasks, and generate reasonable function call sequences and parameter planning accordingly, thereby supporting the scheduling and execution of subsequent algorithm modules. This invention is based on a locally deployed operation and maintenance algorithm service cluster as the execution foundation, managing various operation and maintenance diagnostic algorithms by encapsulating them into microservices. This cluster includes five types of operation and maintenance analysis algorithms: component characteristic correction algorithms, energy efficiency monitoring algorithms, performance degradation analysis algorithms, optimal control trajectory recommendation algorithms, and comprehensive quantitative diagnostic model algorithms. These analysis modules employ a technical approach that combines physical mechanism-based models with machine learning-based data-driven methods. For example, they utilize thermodynamic equilibrium mechanisms and component parameter models to calibrate gas turbine performance, while introducing machine learning methods such as neural networks to predict performance degradation trends, thereby improving the accuracy and reliability of diagnosis. A microservice architecture based on the FastAPI framework is used to implement interface calls and parallel scheduling of various algorithm modules, thus meeting the needs of real-time diagnosis. A function call-based gas turbine fault quantification diagnosis method includes the following steps: S1: Receive a description of gas turbine operation anomalies input by the user in natural language form, perform semantic parsing on the anomaly description, and generate a structured task representation; S2: Based on a pre-built workflow knowledge base, perform matching analysis on the structured task representation to generate a candidate algorithm call sequence. The workflow knowledge base includes at least a gas turbine domain knowledge graph, a tool and property package knowledge base, and a document and historical case knowledge base. S3: Based on the candidate algorithm call sequence, call the locally deployed operation and maintenance algorithm cluster to perform phased calculations. The operation and maintenance algorithm cluster includes component characteristic correction algorithm, energy efficiency monitoring algorithm, performance degradation analysis algorithm, optimal control trajectory recommendation algorithm, and comprehensive quantitative diagnostic model algorithm. S4: After each algorithm call, the calculation result is fed back to the large language model to update the current task state, and the decision is made on whether to continue executing subsequent algorithm calls based on the updated state; S5: When the diagnostic termination conditions are met, a comprehensive analysis of the calculation results of each stage is performed to generate a structured diagnostic report that includes fault probability ranking, key performance index deviation values, and optimization suggestions.
[0007] Furthermore, S1 includes the following steps: S11: Collect gas turbine performance test reports, historical maintenance records, analysis documents, and operation and maintenance logs to build a corpus in the gas turbine field; S12: Perform domain adaptation fine-tuning and alignment on the general large language model to form a vertical domain large model for gas turbine operation and maintenance, enabling it to accurately understand the names of gas turbine components, operating parameters, typical fault phenomena, and engineering expressions; supervised training is conducted using an instruction-response format, so that the model learns to map "natural language fault description" to "structured task representation", using the expression of equation (1): , in, A structured task description for gas turbine diagnostic tasks. Provide users with fault descriptions and analysis requirements input in natural language. This is a knowledge graph for the field of gas turbine operation and maintenance. For dialogue with history, This is a finely tuned large-scale model for the gas turbine operation and maintenance field. Through this step, the model acquires semantic parsing capabilities within this field, enabling it to transform user-input natural language descriptions into structured descriptions corresponding to the analysis tasks.
[0008] Furthermore, S2 specifically includes the following steps: S21: A multi-level workflow knowledge base is pre-built, including a gas turbine domain knowledge graph database, a tool and property package knowledge base, and a document and historical case knowledge base. The gas turbine domain knowledge graph database constructs structured graph data of components, failure modes, anomalies, and causal relationships. The tool and property package knowledge base abstracts all callable algorithm modules into function description units, with each function containing input parameter definitions, output result descriptions, applicable scenario tags, and calling examples. The document and historical case knowledge base encodes unit technical manuals and historical diagnostic reports, constructing a retrieval-enhanced database to supplement information in complex scenarios. S22: Set up an intelligent agent collaborative operation mechanism; at the system architecture level, the overall operation logic is divided into four functional modules: task understanding and input analysis, tool selection and execution strategy formulation, memory and knowledge management, and action execution and result optimization; when the user inputs a natural language description, the task understanding module first converts the text into a structured task representation and locates relevant entities and fault paths in the knowledge graph; then the tool selection module combines the tool knowledge base and historical memory information to generate a candidate algorithm call sequence and execution order, using the expression (3): , in, This represents the gas turbine operation and maintenance algorithm module. This represents the set of gas turbine operation and maintenance algorithm services in the system. This indicates the currently selected algorithm module that best matches the user's gas turbine operation and maintenance problem analysis needs and existing input parameters. This indicates that the algorithm is invoked under the constraints of the current task and operational knowledge. The probability; the large language model selects the most suitable operation and maintenance analysis algorithm from the algorithm service set based on the structured task and knowledge graph information.
[0009] Furthermore, in S3, a cluster of gas turbine operation and maintenance algorithms is developed, including component characteristic correction algorithms, energy efficiency monitoring algorithms, performance degradation analysis algorithms, optimal control trajectory recommendation algorithms, and comprehensive quantitative diagnostic model algorithms. The component characteristic correction algorithm uses the baseline estimation method to correct the characteristic lines of gas turbine components; the energy efficiency monitoring algorithm uses the characteristic line method to model the gas turbine body and uses the corrected model as a reference point to monitor the energy efficiency of each component; after detecting equipment performance degradation, the performance degradation analysis algorithm is used to correct the characteristic lines; a performance degradation factor is introduced into the heavy-duty gas turbine component model to convert the gas turbine physical model into a performance degradation diagnostic model, and further comprehensive diagnosis of gas turbine failure modes is carried out. Based on the degradation factor obtained by the algorithm, it is determined whether the component performance is approaching a certain failure state; after determining the degradation state of the component, the optimal control trajectory recommendation algorithm is called to determine the new control point based on the historical operating conditions of the component; finally, the degradation model algorithm is used to schedule the maintenance of degradation models for different failure modes. The above analysis formulas based on the gas turbine operating mechanism are all written in the corresponding algorithm functions and are triggered by the large language model based on the prompt words input by the user; all algorithms are independently developed as modular functions, defining corresponding prompt words and encapsulating them in subsequent steps.
[0010] Furthermore, in S4, the specific steps include: S41: The FastAPI framework is used to build an HTTP REST API for each algorithm module, and it is deployed locally through Docker containers to achieve standardized algorithm calls; all algorithm interfaces are registered to the function list during system initialization for the large language model to call during inference, and its API calls adopt the expression (2): , in, This represents the address of each gas turbine operation and maintenance algorithm service interface. Indicates the parameters to be called; S42: Deploy the large language model service and implement function call integration; deploy the domain-aligned model as a local inference service, enable the function call interface, so that the model can output structured call instructions during the inference process; the system adopts a cyclic execution process: the model first generates a function based on the current context, and the call decision adopts the expression (4): , in, This represents the call chain of the gas turbine operation and maintenance algorithm generated by the system. Indicates the first In order to analyze the failure mode of a gas turbine, the function call first needs to obtain the operating parameters of the healthy state, call the component characteristic correction algorithm to obtain the health state characteristics of the equipment, then use the current operating parameters of the equipment to call the performance degradation analysis algorithm to solve the degradation factor, and then match it with the most similar failure mode features. Then each function call unit adopts the expression (5): , in, For the gas turbine operation and maintenance algorithm module being called, The input parameters for this algorithm are used; the system calls the corresponding algorithm service and obtains the result using expression (6): , in, Indicates the first The output results of each operation and maintenance algorithm module are fed back to the model as new context input. The model determines whether to continue calling other algorithms based on the updated state until the diagnostic termination condition is met.
[0011] Furthermore, in S5, after the system's diagnostic process is completed, the calculation results of each stage are integrated and merged to output a structured diagnostic report, including fault mode probability ranking, key performance index deviation values, health index change trends, and recommended control or maintenance suggestions; at the same time, the function call paths and their explanations are retained; the final report is generated in the form of structured data.
[0012] This invention also provides a gas turbine fault quantification diagnosis system based on function calls, comprising: The user input module is used to receive natural language descriptions of gas turbine malfunctions. The large language model scheduling module is used to perform semantic parsing on the input content and generate an algorithm call sequence. After each algorithm execution, the large language model scheduling module receives feedback results and dynamically determines subsequent algorithm calls based on the updated task status. The large language model scheduling module supports a loop function call mechanism, continuously executing the iterative process of "generating call instructions - executing algorithms - providing feedback results - updating status" until the termination condition is met. The workflow knowledge base module is used to provide the large language model with knowledge graphs, tool knowledge bases, and document knowledge bases in this field. The operation and maintenance algorithm service module is used to execute the algorithm calculations in the call sequence; The results generation module is used to generate a structured diagnostic report by comprehensively calculating the results.
[0013] Furthermore, the operation and maintenance algorithm service module includes a component characteristic correction submodule, an energy efficiency monitoring submodule, a performance degradation analysis submodule, an optimal control trajectory recommendation submodule, and a comprehensive quantitative diagnosis submodule.
[0014] Furthermore, the large language model scheduling module includes a task understanding and input analysis module, a tool selection and execution strategy formulation module, a memory and knowledge management module, and an action execution and result optimization module.
[0015] Furthermore, the results generation module retains the function call path and intermediate calculation results to provide traceable diagnostic evidence.
[0016] Compared with existing technologies, this invention introduces a function call mechanism to deeply integrate a large language model with an engineering-verified gas turbine operation and maintenance algorithm cluster. This allows the large model to serve as the core of intelligent scheduling and process planning, driving specialized algorithms to complete quantitative calculations and analyses, thereby significantly improving the engineering reliability and numerical accuracy of diagnostic results. General-purpose large models are prone to inference bias or unexplainable results when lacking professional constraints. This invention addresses this by constructing a knowledge graph, tool knowledge base, and workflow knowledge base in the gas turbine field to structurally constrain the model's inference path. This ensures that the order of calls, parameter generation, and algorithm selection are all constrained by engineering knowledge and physical mechanisms, resulting in more accurate and interpretable outputs. By constructing a service-oriented deployment of operation and maintenance algorithm clusters, multi-algorithm collaborative computing and phased quantitative analysis are achieved. This allows for the dynamic combination of algorithm modules such as component characteristic correction, performance degradation assessment, energy efficiency monitoring, and comprehensive quantitative diagnostic models according to different fault scenarios, forming a scalable modular diagnostic system. This improves the system's adaptability to complex coupled faults. Furthermore, by introducing a memory mechanism and dynamic path planning mechanism through an intelligent agent architecture, the system can learn from historical diagnostic experience and continuously optimize algorithm calling strategies and execution processes, achieving an improvement from "static rule judgment" to "evolvable intelligent decision-making." Attached Figure Description
[0017] Figure 1 This is a schematic diagram illustrating the implementation steps of a gas turbine fault quantification diagnosis method and system based on function calls proposed in this invention. Figure 2 This is a schematic diagram of the overall architecture of a gas turbine fault quantification diagnosis method and system based on function calls according to the present invention. Figure 3 This is a schematic diagram of the intelligent agent function collaborative operation mechanism of a gas turbine fault quantification diagnosis method and system based on function calls according to the present invention. Figure 4 This is a schematic diagram of the function call timing flow of a gas turbine fault quantification diagnosis method and system based on function calls according to the present invention. Figure 5 This is a schematic diagram of the containerized deployment structure of the algorithm service cluster of a gas turbine fault quantification diagnosis method and system based on function calls according to the present invention. Figure 6 This is a schematic diagram of a gas turbine fault quantification diagnosis system based on function calls according to the present invention. Detailed Implementation
[0018] like Figures 1 to 5 As shown, a gas turbine fault quantification diagnosis method based on function calls includes the following steps: S1: Receive a description of gas turbine operation anomalies input by the user in natural language form, perform semantic parsing on the anomaly description, and generate a structured task representation; S2: Based on a pre-built workflow knowledge base, perform matching analysis on the structured task representation to generate a candidate algorithm call sequence. The workflow knowledge base includes at least a gas turbine domain knowledge graph, a tool and property package knowledge base, and a document and historical case knowledge base. S3: Based on the candidate algorithm call sequence, call the locally deployed operation and maintenance algorithm cluster to perform phased calculations. The operation and maintenance algorithm cluster includes component characteristic correction algorithm, energy efficiency monitoring algorithm, performance degradation analysis algorithm, optimal control trajectory recommendation algorithm, and comprehensive quantitative diagnostic model algorithm. S4: After each algorithm call, the calculation result is fed back to the large language model to update the current task state, and the decision is made on whether to continue executing subsequent algorithm calls based on the updated state; S5: When the diagnostic termination conditions are met, a comprehensive analysis of the calculation results of each stage is performed to generate a structured diagnostic report that includes fault probability ranking, key performance index deviation values, and optimization suggestions.
[0019] In a preferred embodiment, S1 includes the following steps: S11: Collect gas turbine performance test reports, historical maintenance records, analysis documents, and operation and maintenance logs to build a corpus in the gas turbine field; S12: Perform domain adaptation fine-tuning and alignment on the general large language model to form a vertical domain large model for gas turbine operation and maintenance, enabling it to accurately understand the names of gas turbine components, operating parameters, typical fault phenomena, and engineering expression methods; based on fault samples and engineering documents in the gas turbine operation and maintenance field, incrementally train the pre-trained large language model through low-rank adaptation LoRA, so that the model can acquire professional knowledge in the gas turbine field and improve its language understanding and reasoning ability for related fault phenomena; supervised training is carried out using the instruction-response form, so that the model learns to map "natural language fault description" to "structured task representation", using the expression of equation (1): , in, A structured task description for gas turbine diagnostic tasks. Provide users with fault descriptions and analysis requirements input in natural language. This is a knowledge graph for the field of gas turbine operation and maintenance. For dialogue with history, This is a finely tuned large-scale model for the gas turbine operation and maintenance field. Through this step, the model acquires semantic parsing capabilities within this field, enabling it to transform user-input natural language descriptions into structured descriptions corresponding to the analysis tasks.
[0020] In a preferred embodiment, step S2 specifically includes the following steps: S21: A multi-level workflow knowledge base is pre-built, including a gas turbine domain knowledge graph database, a tool and property package knowledge base, and a document and historical case knowledge base. The gas turbine domain knowledge graph database constructs structured graph data of components, failure modes, anomalies, and causal relationships. The tool and property package knowledge base abstracts all callable algorithm modules into function description units, each function containing input parameter definitions, output result descriptions, applicable scenario tags, and calling examples. The document and historical case knowledge base encodes unit technical manuals and historical diagnostic reports, constructing a retrieval enhancement database to supplement information in complex scenarios. Entities and relationships in the knowledge graph provide structured knowledge support for the trained large language model reasoning, enabling the system to retrieve relevant knowledge based on user natural language input and output diagnostic conclusions. S22: Set up an intelligent agent collaborative operation mechanism; at the system architecture level, the overall operation logic is divided into four functional modules: task understanding and input analysis, tool selection and execution strategy formulation, memory and knowledge management, action execution and result optimization. The task understanding module first performs semantic analysis on the natural language input of the operation and maintenance personnel to identify the equipment components, operating status and abnormal characteristics involved; the tool selection and strategy module selects appropriate diagnostic functions from the encapsulated algorithm services and generates an execution plan based on the task objectives and context; the memory and knowledge management module continuously stores the context information, knowledge graph content and historical results in the diagnostic process to form short-term and long-term memory, providing retrieval and reference for the model; the action execution and result optimization module is responsible for calling the selected algorithm to perform fault reasoning calculation and optimizing the diagnostic scheme based on the calculation results; each module closely cooperates through function calls and data interaction to jointly form a closed-loop intelligent scheduling process; when the user inputs a natural language description, the task understanding module first converts the text into a structured task representation and locates the relevant entities and fault paths in the knowledge graph; then the tool selection module combines the tool knowledge base and historical memory information to generate a candidate algorithm call sequence and execution order, using the expression (3): , in, This represents the gas turbine operation and maintenance algorithm module. This represents the set of gas turbine operation and maintenance algorithm services in the system. This indicates the currently selected algorithm module that best matches the user's gas turbine operation and maintenance problem analysis needs and existing input parameters. This indicates that the algorithm is invoked under the constraints of the current task and operational knowledge. The probability; the large language model selects the most suitable operation and maintenance analysis algorithm from the algorithm service set based on the structured task and knowledge graph information.
[0021] As a preferred embodiment, in S3, a gas turbine operation and maintenance algorithm cluster is developed, including component characteristic correction algorithm, energy efficiency monitoring algorithm, performance degradation analysis algorithm, optimal control trajectory recommendation algorithm, and comprehensive quantitative diagnostic model algorithm; The component characteristic correction algorithm uses the baseline estimation method to correct the characteristic lines of gas turbine components; the energy efficiency monitoring algorithm uses the characteristic line method to model the gas turbine body and uses the corrected model as a reference point to monitor the energy efficiency of each component; after detecting equipment performance degradation, the performance degradation analysis algorithm is used to correct the characteristic lines; a performance degradation factor is introduced into the heavy-duty gas turbine component model to convert the gas turbine physical model into a performance degradation diagnostic model, and further comprehensive diagnosis of gas turbine failure modes is carried out. Based on the degradation factor obtained by the algorithm, it is determined whether the component performance is approaching a certain failure state; after determining the degradation state of the component, the optimal control trajectory recommendation algorithm is called to determine the new control point based on the historical operating conditions of the component; finally, the degradation model algorithm is used to schedule the maintenance of degradation models for different failure modes. The above analysis formulas based on the gas turbine operating mechanism are all written in the corresponding algorithm functions and are triggered by the large language model based on the prompt words input by the user; all algorithms are independently developed as modular functions, defining corresponding prompt words and encapsulating them in subsequent steps.
[0022] In a preferred embodiment, step S4 specifically includes: S41: The FastAPI framework is used to build an HTTP REST API for each algorithm module, and it is deployed locally through Docker containers to achieve standardized algorithm calls; the FastAPI framework is used to develop RESTful interfaces for each operation and maintenance algorithm module, and they are packaged into Docker containers for operation to achieve environment isolation and resource management; each modular service provides calling capabilities to the outside world through API interfaces, so that the large language model scheduling engine can access the required algorithms through function calls; all algorithm interfaces are registered to the function list during system initialization for the large language model to call during inference, and its API calls adopt the expression (2): , in, This represents the address of each gas turbine operation and maintenance algorithm service interface. Indicates the parameters to be called; S42: Deploy the large language model service and implement function call integration; deploy the domain-aligned model as a local inference service, enable the function call interface, so that the model can output structured call instructions during the inference process; the system adopts a cyclic execution process: the model first generates a function based on the current context, and the call decision adopts the expression (4): , in, This represents the call chain of the gas turbine operation and maintenance algorithm generated by the system. Indicates the first In order to analyze the failure mode of a gas turbine, the function call first needs to obtain the operating parameters of the healthy state, call the component characteristic correction algorithm to obtain the health state characteristics of the equipment, then use the current operating parameters of the equipment to call the performance degradation analysis algorithm to solve the degradation factor, and then match it with the most similar failure mode features. Then each function call unit adopts the expression (5): , in, For the gas turbine operation and maintenance algorithm module being called, The input parameters for this algorithm are used; the system calls the corresponding algorithm service and obtains the result using expression (6): , in, Indicates the first The output results of each operation and maintenance algorithm module are fed back to the model as new context input. The model determines whether to continue calling other algorithms based on the updated state until the diagnostic termination condition is met.
[0023] In a preferred embodiment, in S5, after the system's diagnostic process is completed, the calculation results of each stage are integrated and output as a structured diagnostic report, including fault mode probability ranking, key performance index deviation values, health index change trends, and recommended control or maintenance suggestions. At the same time, the function call path and the explanation of the basis are retained, so that the diagnostic process is traceable and engineering interpretable. The final report is generated in the form of structured data and can be converted into a visual display interface for operation and maintenance personnel to use.
[0024] like Figure 4 As shown, the present invention also provides a gas turbine fault quantification diagnosis system based on function calls, comprising: The user input module is used to receive natural language descriptions of gas turbine malfunctions. The large language model scheduling module performs semantic parsing on the input content and generates an algorithm call sequence. This semantic parsing includes mapping natural language anomaly descriptions to equipment component entities, operating parameter entities, and anomaly feature entities, and performing entity association and localization within the gas turbine domain knowledge graph. After each algorithm execution, the large language model scheduling module receives feedback results and dynamically determines subsequent algorithm calls based on the updated task status. The module supports a cyclic function call mechanism, continuously executing an iterative process of "generating call instructions—executing algorithms—providing feedback results—updating status" until a termination condition is met. The large language model scheduling module includes task understanding and input analysis. The system comprises three modules: a tool selection and execution strategy formulation module, a memory and knowledge management module, and an action execution and result optimization module. The memory and knowledge management module stores historical diagnostic task data and optimizes subsequent algorithm call paths, providing strategy references for similar tasks. It continuously records intermediate results and decision paths during the current diagnostic process and provides strategy references for similar task scenarios. The action execution and result optimization module actually triggers algorithm services based on the planned results and performs outlier detection on the returned data. If an anomaly is found, it feeds back to the preceding modules to replan the call path. Through this mechanism, a dynamic closed-loop optimization process from task understanding to result verification is achieved. The workflow knowledge base module is used to provide the large language model with knowledge graphs, tool knowledge bases and file knowledge bases in this field; the tool and property package knowledge base includes function names, input parameter definitions, output parameter descriptions and applicable scenario tags, and the large language model generates function call instructions based on the tags; The operation and maintenance algorithm service module is used to execute the algorithm calculations in the call sequence. The operation and maintenance algorithm service module includes a component characteristic correction submodule, an energy efficiency monitoring submodule, a performance degradation analysis submodule, an optimal control trajectory recommendation submodule, and a comprehensive quantitative diagnosis submodule. The operation and maintenance algorithm cluster builds standardized interfaces through the FastAPI framework and encapsulates and deploys them through Docker containers to realize modular service calls. Docker-based containerized deployment ensures that the operating environments of each algorithm are isolated from each other, improving system stability and maintainability. The results generation module is used to generate a structured diagnostic report by comprehensively calculating the results; the results generation module retains the function call path and intermediate calculation results to provide traceable diagnostic basis.
[0025] This invention, through a function-call-based gas turbine fault quantification diagnosis system and corresponding diagnostic method, constructs a function-call-based intelligent agent system for gas turbine fault quantification diagnosis. This system uses a large language model as the scheduling core and a service-oriented operation and maintenance algorithm cluster as the computational core, achieving a complete closed-loop process from natural language input to multi-algorithm collaborative computation and structured report output. Regarding diagnostic accuracy, by combining the semantic understanding capabilities of the large language model with an engineering-verified operation and maintenance algorithm cluster, the uncertainty caused by relying solely on empirical inferences from statistical language models is avoided. Fault determination is based on clear physical mechanism calculations and quantification models, significantly improving fault identification accuracy and false positive suppression capabilities. In terms of quantification capabilities, this method achieves standardized correction of key performance parameters, energy efficiency assessment, and health index analysis through multi-algorithm collaborative computation, elevating the diagnostic results from "qualitative judgment" to "measurable numerical evaluation," improving the consistency of diagnostic results. The complete function call path and intermediate calculation results are preserved, ensuring that each conclusion is traceable to specific algorithm modules and engineering basis, enhancing the interpretability of the results.
[0026] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.
[0027] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A method for quantitative diagnosis of gas turbine faults based on function calls, characterized in that: Includes the following steps: S1: Receive a description of gas turbine operation anomalies input by the user in natural language form, perform semantic parsing on the anomaly description, and generate a structured task representation; S2: Based on a pre-built workflow knowledge base, perform matching analysis on the structured task representation to generate a candidate algorithm call sequence. The workflow knowledge base includes at least a gas turbine domain knowledge graph, a tool and property package knowledge base, and a document and historical case knowledge base. S3: Based on the candidate algorithm call sequence, call the locally deployed operation and maintenance algorithm cluster to perform phased calculations. The operation and maintenance algorithm cluster includes component characteristic correction algorithm, energy efficiency monitoring algorithm, performance degradation analysis algorithm, optimal control trajectory recommendation algorithm, and comprehensive quantitative diagnostic model algorithm. S4: After each algorithm call, the calculation result is fed back to the large language model to update the current task state, and the decision is made on whether to continue executing subsequent algorithm calls based on the updated state; S5: When the diagnostic termination conditions are met, a comprehensive analysis of the calculation results of each stage is performed to generate a structured diagnostic report that includes fault probability ranking, key performance index deviation values, and optimization suggestions.
2. The gas turbine fault quantification diagnosis method based on function calls according to claim 1, characterized in that: S1 includes the following steps: S11: Collect gas turbine performance test reports, historical maintenance records, analysis documents, and operation and maintenance logs to build a corpus in the gas turbine field; S12: Perform domain adaptation fine-tuning and alignment on the general large language model to form a vertical domain large model for gas turbine operation and maintenance, enabling it to accurately understand the names of gas turbine components, operating parameters, typical fault phenomena, and engineering expressions; supervised training is conducted using an instruction-response format, so that the model learns to map "natural language fault description" to "structured task representation", using the expression of equation (1): ， in, A structured task description for gas turbine diagnostic tasks. Provide users with fault descriptions and analysis requirements input in natural language. This is a knowledge graph for the field of gas turbine operation and maintenance. For dialogue with history, This is a finely tuned large-scale model for the gas turbine operation and maintenance field. Through this step, the model acquires semantic parsing capabilities within this field, enabling it to transform user-input natural language descriptions into structured descriptions corresponding to the analysis tasks.
3. The gas turbine fault quantification diagnosis method based on function calls according to claim 1, characterized in that: S2 specifically includes the following steps: S21: A multi-level workflow knowledge base is pre-built, including a gas turbine domain knowledge graph database, a tool and property package knowledge base, and a document and historical case knowledge base. The gas turbine domain knowledge graph database constructs structured graph data of components, failure modes, anomalies, and causal relationships. The tool and property package knowledge base abstracts all callable algorithm modules into function description units, with each function containing input parameter definitions, output result descriptions, applicable scenario tags, and calling examples. The document and historical case knowledge base encodes unit technical manuals and historical diagnostic reports, constructing a retrieval-enhanced database to supplement information in complex scenarios. S22: Set up an intelligent agent collaborative operation mechanism; at the system architecture level, the overall operation logic is divided into four functional modules: task understanding and input analysis, tool selection and execution strategy formulation, memory and knowledge management, and action execution and result optimization; when the user inputs a natural language description, the task understanding module first converts the text into a structured task representation and locates relevant entities and fault paths in the knowledge graph; then the tool selection module combines the tool knowledge base and historical memory information to generate a candidate algorithm call sequence and execution order, using the expression (3): , in, This represents the gas turbine operation and maintenance algorithm module. This represents the set of gas turbine operation and maintenance algorithm services in the system. This indicates the currently selected algorithm module that best matches the user's gas turbine operation and maintenance problem analysis needs and existing input parameters. This indicates that the algorithm is invoked under the constraints of the current task and operational knowledge. The probability; the large language model selects the most suitable operation and maintenance analysis algorithm from the algorithm service set based on the structured task and knowledge graph information.
4. The gas turbine fault quantification diagnosis method based on function calls according to claim 1, characterized in that: In S3, a cluster of gas turbine operation and maintenance algorithms was developed, including component characteristic correction algorithms, energy efficiency monitoring algorithms, performance degradation analysis algorithms, optimal control trajectory recommendation algorithms, and comprehensive quantitative diagnostic model algorithms. The component characteristic correction algorithm uses the baseline estimation method to correct the characteristic lines of gas turbine components; the energy efficiency monitoring algorithm uses the characteristic line method to model the gas turbine body and uses the corrected model as a reference point to monitor the energy efficiency of each component; after detecting equipment performance degradation, the performance degradation analysis algorithm is used to correct the characteristic lines; a performance degradation factor is introduced into the heavy-duty gas turbine component model to convert the gas turbine physical model into a performance degradation diagnostic model, and further comprehensive diagnosis of gas turbine failure modes is carried out. Based on the degradation factor obtained by the algorithm, it is determined whether the component performance is approaching a certain failure state; after determining the degradation state of the component, the optimal control trajectory recommendation algorithm is called to determine the new control point based on the historical operating conditions of the component; finally, the degradation model algorithm is used to schedule the maintenance of degradation models for different failure modes. The above analysis formulas based on the gas turbine operating mechanism are all written in the corresponding algorithm functions and are triggered by the large language model based on the prompt words input by the user; all algorithms are independently developed as modular functions, defining corresponding prompt words and encapsulating them in subsequent steps.
5. The gas turbine fault quantification diagnosis method based on function calls according to claim 4, characterized in that: In S4, the specific steps include: S41: The FastAPI framework is used to build an HTTP REST API for each algorithm module, and it is deployed locally through Docker containers to achieve standardized algorithm calls; all algorithm interfaces are registered to the function list during system initialization for the large language model to call during inference, and its API calls adopt the expression (2): ， in, This represents the address of each gas turbine operation and maintenance algorithm service interface. Indicates the parameters to be called; S42: Deploy the large language model service and implement function call integration; deploy the domain-aligned model as a local inference service, enable the function call interface, so that the model can output structured call instructions during the inference process; the system adopts a cyclic execution process: the model first generates a function based on the current context, and the call decision adopts the expression (4): ， in, This represents the call chain of the gas turbine operation and maintenance algorithm generated by the system. Indicates the first In order to analyze the failure mode of a gas turbine, the function call first needs to obtain the operating parameters of the healthy state, call the component characteristic correction algorithm to obtain the health state characteristics of the equipment, then use the current operating parameters of the equipment to call the performance degradation analysis algorithm to solve the degradation factor, and then match it with the most similar failure mode features. Then each function call unit adopts the expression (5): , in, For the gas turbine operation and maintenance algorithm module being called, The input parameters for this algorithm are used; the system calls the corresponding algorithm service and obtains the result using expression (6): ， in, Indicates the first The output results of each operation and maintenance algorithm module are fed back to the model as new context input. The model determines whether to continue calling other algorithms based on the updated state until the diagnostic termination condition is met.
6. The gas turbine fault quantification diagnosis method based on function calls according to claim 1, characterized in that: In S5, after the system's diagnostic process is completed, the calculation results of each stage are integrated and merged to output a structured diagnostic report that includes fault mode probability ranking, key performance index deviation values, health index change trends, and recommended control or maintenance suggestions. The function call path and its supporting documentation are also retained; the final report is generated in structured data format.
7. A gas turbine fault quantification diagnostic system based on function calls, used to execute the diagnostic method according to any one of claims 1 to 6, characterized in that: include: The user input module is used to receive natural language descriptions of gas turbine malfunctions. The large language model scheduling module is used to perform semantic parsing of the input content and generate an algorithm call sequence; The large language model scheduling module receives feedback results after each algorithm execution and dynamically determines subsequent algorithm calls based on the updated task status. The large language model scheduling module supports a cyclic function call mechanism, continuously executing the iterative process of "generating call instructions - executing algorithms - providing feedback results - updating status" until the termination condition is met. The workflow knowledge base module is used to provide the large language model with knowledge graphs, tool knowledge bases, and document knowledge bases in this field. The operation and maintenance algorithm service module is used to execute the algorithm calculations in the call sequence; The results generation module is used to generate a structured diagnostic report by comprehensively calculating the results.
8. The gas turbine fault quantification diagnosis system based on function calls according to claim 7, characterized in that: The operation and maintenance algorithm service module includes a component characteristic correction submodule, an energy efficiency monitoring submodule, a performance degradation analysis submodule, an optimal control trajectory recommendation submodule, and a comprehensive quantitative diagnosis submodule.
9. A gas turbine fault quantification diagnosis system based on function calls according to claim 7, characterized in that: The large language model scheduling module includes a task understanding and input analysis module, a tool selection and execution strategy formulation module, a memory and knowledge management module, and an action execution and result optimization module.
10. A gas turbine fault quantification diagnosis system based on function calls according to claim 7, characterized in that: The results generation module retains the function call path and intermediate calculation results to provide traceable diagnostic evidence.