An automated fault diagnosis method for industrial systems based on a large language model

Abstract

The present invention discloses an automated fault diagnosis method for industrial systems based on a large language model. The method constructs a three-layer mapping system of industrial data, natural language description, and knowledge reasoning, semantically converts on-site multi-source sensor data, and constructs a quantitative calculation model based on a large language model based on historical data and logs. With the help of retrieval-enhanced generation technology, fault cases are matched in real time as a reference, and input into a knowledge reasoning model based on a large language model together with abnormal information to generate a structured logical reasoning chain, and further output a diagnostic conclusion containing candidate faults, cause analysis, and disposal suggestions. At the same time, through user feedback and reinforcement learning mechanisms, the model and knowledge base are adaptively updated, effectively making up for the shortcomings of traditional static rules and expert experience, significantly improving the accuracy, interpretability, and robustness of fault detection, and adapting to complex and changing working conditions.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial intelligent diagnosis, and in particular relates to an automated fault diagnosis method based on a large language model. Background Art

[0002] In the field of industrial fault diagnosis, traditional expert systems primarily rely on domain experts to manually write and maintain rule bases. For example, they use logic such as "abnormally high parameter A + abnormally low parameter B → fault X" to make inferences. However, when scaling to multiple scenarios and multi-variable operating conditions, such systems often face bottlenecks such as high rule base maintenance costs, insufficient logic coverage, and difficulty adapting to dynamic operating conditions. Furthermore, judgment methods based on human experience are subject to significant subjectivity in practice: operators or engineers may provide different diagnostic conclusions for the same fault phenomenon, making it difficult to ensure consistent and accurate results. With the continuous increase in the scale of industrial systems and the number of monitored variables, fault diagnosis methods that rely on individual experience are increasingly unable to meet the needs of complex scenarios. The rapid development of Large Language Model (LLM) technology, with its knowledge base and logical reasoning capabilities, built from extensive literature and case studies, has provided a new approach to fault diagnosis. However, the lack of a unified mapping between sensor encoding and calculation parameters and natural language descriptions in industrial systems makes it difficult for LLMs to intuitively understand multi-source data. Furthermore, traditional static thresholds cannot accurately quantify the degree of parameter anomalies. There is an urgent need to establish dynamic normal ranges based on historical data, process adjustments, and environmental factors. Furthermore, LLM relies solely on inherent, static knowledge and may not be able to absorb new fault information and cases in a timely manner during long-term operation, potentially leading to misdiagnoses or missed diagnosis. Therefore, effectively combining industrial data with the semantic reasoning capabilities of LLM to form an intelligent diagnostic framework that can be rapidly updated and dynamically adapted to different operating conditions has become a pressing technical challenge. Summary of the Invention

[0003] The present invention proposes an automated diagnosis method based on a three-layer mapping of industrial data, natural language, and knowledge reasoning. The method constructs an industrial system operation data set and log library, establishes a two-way mapping between industrial variables and natural language, and uses a quantitative calculation model to dynamically construct the normal range of variables and output abnormal information. With the help of Retrieval-Augmented Generation (RAG) technology, fault cases are retrieved. Based on the abnormal information and the above cases, the knowledge reasoning model generates a complete logical reasoning chain, thereby achieving accurate industrial system fault diagnosis. At the same time, user feedback and reinforcement learning algorithms are combined to realize the adaptive update of the knowledge reasoning model, and new fault cases and fault modes are regularly added to the domain knowledge base to ensure that when dealing with variable working conditions and new faults, high precision, real-time response and adaptive update capabilities are always available. In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0004] An automated fault diagnosis method for industrial systems based on a large language model, comprising the steps of:

[0005] S1, constructs the industrial system operation data set and preprocesses the data in the industrial system operation data set;

[0006] S2, builds semantic mapping relationships between industrial variables and converts industrial data into natural language descriptions;

[0007] S3, builds a quantitative computing model based on a large language model, outputs abnormal content based on semantic fusion and comprehensive reasoning;

[0008] S4, builds a knowledge reasoning model based on a large language model to diagnose industrial system faults;

[0009] S5, optimizing the knowledge reasoning model based on the expert feedback mechanism, and adaptively updating the knowledge base and the knowledge reasoning model;

[0010] Step S4 constructs a knowledge reasoning model based on a large language model to implement fault diagnosis of industrial systems, including:

[0011] Step S41: Based on the retrieval-enhancement-generation technology, fault cases and diagnostic experiences that are highly relevant to the current abnormal state are retrieved in real time from the industrial system domain knowledge base constructed in step S14;

[0012] Step S42: Based on the large language model, a comprehensive analysis is performed on the intrinsic correlation between the abnormal state and historical fault cases, a logical reasoning chain is generated, and a fault diagnosis conclusion is output;

[0013] Step S41 includes: using LLaMA word vector encoder Convert the data to be retrieved and each document into word vectors and :

[0014] ;

[0015] Use cosine similarity to calculate the similarity score between the data to be retrieved and each document ,in Represents the point operation, Represents the modulus of a vector:

[0016] ;

[0017] Select the one with the highest similarity Documents constitute a collection :

[0018] ,

[0019] The exception information and The cases are concatenated and used as input for the large language model. :

[0020] ;

[0021] Step S5 optimizes the knowledge reasoning model based on the expert feedback mechanism, and adaptively updates the knowledge base and the knowledge reasoning model, including:

[0022] Step S51, feedback optimization: establish a structured user feedback collection mechanism, record the operator's evaluation and scoring of the diagnosis conclusion, use the feedback data as input to perform reinforcement learning through the GRPO algorithm, and optimize the strategy of the knowledge reasoning model;

[0023] Step S52, iterative update of knowledge base cases: based on expert feedback and the latest industry literature, equipment maintenance manuals and expert reports, new failure cases and failure modes are regularly added to the domain knowledge base.

[0024] Furthermore, step S1 includes:

[0025] Step S11: Obtain historical operating data of key equipment in the industrial system, external environmental factors at the corresponding time, physical parameters and load characteristics; construct an industrial system data set ,

[0026] ,in Represents the measured data of the i-th feature, 1≤i≤ , represents the field of real numbers;

[0027] Step S12: pre-process the data in the industrial system dataset X based on 3 The criterion for each column of features Conduct preliminary screening, including is the feature mean of the column, is the characteristic standard deviation of the column, and the observations that meet the range are recorded as : ;

[0028] Based on the industrial business scope, the observations that meet the business scope are recorded as : , the rated operating range of the equipment is ,

[0029] The obtained data set Expressed as:

[0030] Step S13, building an industrial system log library: the system collects raw log data from the production scheduling system, including process parameter change records, scheduling decisions, equipment status updates, and operator notes, and cleans, formats, and structures the log data.

[0031] Step S14, building an industrial system domain knowledge base: sorting and collecting industrial system failure cases from public documents, industry reports, technical manuals, and historical failure logs, standardizing and storing the collected cases in a structured manner to form an industrial system domain knowledge base with multi-dimensional information such as time stamps, failure types, causes, diagnosis and treatment solutions.

[0032] Furthermore, step S2 includes:

[0033] Based on the multi-source monitoring variables obtained in step S11, an industrial variable-natural language description mapping table is constructed, the variables are matched with their standardized natural language descriptions, and a mapping relationship between the industrial variables and the standardized natural language is established.

[0034] Furthermore, step S3 constructs a quantitative calculation model based on the large language model to achieve abnormal content output, including:

[0035] Step S31: construct a quantitative calculation model based on the large language model, generate a dynamic normal range for each variable, and preset reasonable boundaries of the variables;

[0036] In step S32 , based on the dynamic normal range of the variable generated in step S31 , the real-time monitoring data is compared with the preset reasonable boundaries, and the specific values ​​of the parameters exceeding the normal range are output.

[0037] Furthermore, a structured output mechanism is adopted to output the fault diagnosis conclusion including:

[0038] Candidate fault set: The candidate fault set includes possible fault types and their confidence scores based on reasoning chain matching;

[0039] Fault cause analysis: Fault cause analysis involves showing the logical relationship between abnormal parameters and revealing how they lead to specific faults;

[0040] Recommended treatment plan: Based on the recommended treatment plan and case studies, targeted maintenance suggestions are made.

[0041] Furthermore, step S51 includes:

[0042] The GRPO algorithm starts with the old strategy in an update. For the same exception information Sampling out Diagnosis results , compare the evaluation scores of the diagnostic conclusions within the group, and calculate the normalized advantage value of each answer relative to the group mean , on this basis, the objective function of GRPO is written as:

[0043] ,

[0044] in For the new and old strategies The probability ratio of tokens, is the advantage value obtained by scoring the responses within the group and normalizing them, with min Together, we limit the update range. is the hyperparameter of the shear operation, It is control The coefficient of the divergence constraint strength, It is a reference strategy.

[0045] An industrial system automated fault diagnosis system based on a large language model continuously collects data of key system equipment in the industrial system as input, and outputs corresponding diagnosis results based on an industrial system automated fault diagnosis method based on a large language model.

[0046] The advantages of the present invention are as follows: First, by constructing a semantic mapping mechanism for industrial variables, multi-source on-site data (such as sensor code TF and calculation parameter O2_Level) are converted into natural language descriptions (such as "furnace outlet temperature" and "air oxygen content") to address the semantic gap between industrial data and language models. Secondly, based on the large language model, semantic fusion of recent historical data and industrial system log libraries is performed to dynamically infer the normal fluctuation range of parameters and compare the deviation amplitude of abnormal data output in real time; at the same time, relying on RAG technology, fault cases that are highly relevant to the current abnormality are retrieved, and structured diagnostic results are generated in combination with logical reasoning. The output includes a candidate fault set, fault cause analysis, and recommended disposal solutions. Finally, based on user feedback results, the large language model is updated through reinforcement learning, and the fault cases and fault modes in the knowledge base are regularly updated, thereby continuously improving the accuracy and robustness of fault detection and diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of an automated fault diagnosis method for industrial systems based on a large language model according to an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of a method for automated fault diagnosis of industrial systems based on a large language model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The technical solutions of the embodiments of the present invention are explained and described below, but the following embodiments are only preferred embodiments of the present invention and are not exhaustive. Based on the embodiments in the embodiments, other embodiments obtained by those skilled in the art without creative work are all within the scope of protection of the present invention.

[0050] Next, combine the Figure 1 , Attachment Figure 2 The present invention is further described with specific embodiments.

[0051] An embodiment of the present invention is an industrial system automatic fault diagnosis method based on a large language model. The construction process and principle can be referred to Figure 1 and Figure 2 As shown, the specific steps include:

[0052] S1, builds an industrial system operation data set, collects industrial system operation data and performs data preprocessing;

[0053] S2, builds semantic mapping relationships between industrial variables and converts industrial data into natural language descriptions;

[0054] S3, builds a quantitative computing model based on a large language model, outputs abnormal content based on semantic fusion and comprehensive reasoning;

[0055] S4, builds a knowledge reasoning model based on a large language model to diagnose industrial system faults;

[0056] S5, optimizing the knowledge reasoning model based on the expert feedback mechanism, and adaptively updating the knowledge base and the knowledge reasoning model.

[0057] In one embodiment of the present invention, step S1 includes the following steps:

[0058] Step S11: The method for constructing the industrial system operation data set is to obtain the historical operation data of each key equipment in the industrial system (such as boilers, pumps, compressors, heat exchangers, etc.), as well as the external environmental factors (such as ambient temperature, humidity, wind speed, etc.), equipment physical parameters (such as material properties, heat transfer coefficient, design rated power), and load characteristics (such as overall system load changes and key parameter fluctuations) at the corresponding time. Organize them according to time series to form a multi-time point industrial system status data set. Assume that in the sampling task, the sampling time length is , a total of Types of characteristic data, then the industrial system data set Denoted as: ,in Represents the measured data of the i-th feature, 1≤i≤ , Represents the field of real numbers.

[0059] Step S12, data preprocessing includes: first, based on 3 The criterion (i.e., under the assumption of normal distribution, values ​​exceeding the mean ± 3 times the standard deviation are considered abnormal) is used for each column of features. Conduct preliminary screening, including is the feature mean of the column, is the characteristic standard deviation of the column, and the observations that meet the range can be recorded as : ,

[0060] At the same time, combined with industrial business knowledge, such as the rated working range of the equipment , further identify potential abnormal data points, and record the observations that meet the business scope as :

[0061] , the final dataset It can be expressed as:

[0062] The local time window mean filling method is used to fill missing values. The specific method is: the data of 10 valid time steps before and after the time point where the missing value is located are selected to calculate the mean as the filling value. If there is still missing data in the window, only valid data is used for calculation to maintain the time continuity of the data and improve the accuracy and robustness of fault diagnosis.

[0063] In step S13, the industrial system log library construction method is as follows: the system collects raw log data from the production scheduling system, covering information such as process parameter change records, scheduling decisions, equipment status updates, and operator notes, and cleans, formats, and structures the collected logs.

[0064] Step S14: The method for constructing the industrial system domain knowledge base is as follows: systematically collecting and organizing industrial system failure cases from public documents, industry reports, technical manuals, and historical failure logs to construct a failure case knowledge base D. Each case The failure cases described cover descriptions, cause analysis, diagnostic methods, and treatment measures for various equipment failures (such as boilers, pumps, compressors, and heat exchangers). They include both failure examples reported in academic research and abnormal events recorded during actual operation. The collected cases are standardized and structured, forming a database with multi-dimensional information such as time stamps, failure types, causes, diagnosis, and treatment plans.

[0065] Step S2 includes the following steps:

[0066] For the multi-source monitoring variables obtained in step S11, a mapping table between industrial variables and natural language descriptions is constructed. Variables are matched with their standardized natural language descriptions. For example, the variable "TF" is mapped to "furnace outlet temperature" and "O2_Level" is mapped to "air oxygen content" to ensure the universality and consistency of variable names. Secondly, for variables with different physical units and numerical dimensions, a normalization expression strategy is developed to reduce the impact of unit constraints on natural language conversion. For example, for pressure variables, they can be uniformly converted to "MPa" units and attached with normal range reference values ​​so that the large language model can accurately understand the physical meaning of the variables.

[0067] Step S3 includes the following steps:

[0068] Step S31, the method for constructing a quantitative calculation model based on a large language model is as follows: construct a quantitative calculation model based on a large language model, realize semantic fusion and comprehensive reasoning of multi-source industrial data, and dynamically generate the "normal range" of each variable. Specifically, first select the most recent Day (such as By combining historical monitoring data (within 90 days) with text data from industrial system log libraries, a unified model is constructed to account for multiple factors, including seasonal load fluctuations, equipment aging, and operational strategy adjustments, inferring the appropriate fluctuation range for the variable. For example, in a steel mill, during the initial equipment commissioning phase, temperature fluctuations occur but are concentrated between 1450–1460°C. However, during the equipment's stable aging phase, due to long-term high loads and rising ambient temperatures, the system automatically expands the normal range to 1455–1475°C. Compared to traditional fixed-range methods, this dynamic range more flexibly captures phased characteristics, effectively reducing false alarms or missed alarms caused by overlooking process adjustments, environmental changes, and other factors, thereby ensuring subsequent fault detection.

[0069] Step S32: abnormality determination method is:

[0070] Based on the dynamic normal range established in step S31, the real-time monitoring data is compared with the preset reasonable boundaries, and the specific value of the parameter exceeding the normal range is output. Specifically, when the actual value of a parameter falls below the lower normal limit or exceeds the upper normal limit, the system automatically calculates the deviation from the boundary (as a percentage), intuitively quantifying the degree of abnormality. Finally, the abnormality and the corresponding variable name are stored in text form, for example, "The furnace outlet temperature is 7% below the normal range and the fuel flow rate is 10% above the upper normal limit."

[0071] Step S4 includes the following steps:

[0072] Step S41, fault case retrieval based on RAG: This step relies on the abnormal information recorded in step S32 ,Using Retrieval-Augmented Generation (RAG) technology, the fault cases and ,diagnosis experiences that are highly relevant to the current abnormal ,state are retrieved in real time from the industrial domain knowledge base D ,constructed in step S13.

[0073] Specifically, first, use LLama's word vector encoder Convert the data to be retrieved and each document into word vectors and : Then, the cosine similarity is used to calculate the similarity score between the data to be retrieved and each document. ,in Represents the point operation, Represents the modulus of a vector: , select the one with the highest similarity Documents constitute a collection : ;

[0074] Then, the abnormal information With this The cases are stitched together as input to the LLM : ;

[0075] Step S42, fault diagnosis conclusion generation: After inputting the knowledge inference model, the model, leveraging its extensive knowledge base and rigorous logical reasoning mechanisms, conducts a comprehensive analysis of the inherent connections between abnormal information and historical failure cases, generating a detailed logical reasoning chain. This chain not only reveals the causal relationships between various abnormal indicators but also enhances the explainability of the failure, facilitating the tracing and verification of key nodes in the failure's evolution.

[0076] For example, when the input information includes data such as "the furnace outlet temperature is 5% higher and the oxygen content is 10% lower," the knowledge reasoning model will combine its own knowledge and similar input cases to generate a logical reasoning chain of "abnormal temperature increase + insufficient oxygen → decreased combustion efficiency → possible coking." At the same time, through prompts, the knowledge reasoning model adopts a structured output mechanism to subdivide the diagnosis results into three core parts: (1) candidate fault set: possible fault types and their confidence scores based on the reasoning chain matching; (2) fault cause analysis: showing the logical relationship between abnormal parameters and revealing how they lead to specific faults; (3) recommended disposal plan: based on the case, targeted maintenance suggestions are proposed. For example, for the diagnosis of "furnace coking risk (confidence 92%)," the recommendation is "adjust the combustion ratio and check the carbon deposition condition of the nozzle." This strategy of integrating cross-modal reasoning and structured output not only significantly improves the accuracy and comprehensiveness of diagnosis, but also enhances the transparency and credibility of the fault analysis process.

[0077] Step S5 includes the following steps:

[0078] Step S51: User Feedback Collection and Proximal Policy Generalized Reward Policy Optimization (GRPO) Feedback Optimization: A structured user feedback collection mechanism is established to record operators' evaluations and scores of diagnostic conclusions. This feedback data is then used for reinforcement learning using the GRPO algorithm to optimize the LLM's reasoning strategy.

[0079] Specifically, the GRPO algorithm starts with the old strategy in an update For the same exception information Sampling out Diagnosis results , and use the operator's evaluation and scoring of the diagnostic conclusion to make intra-group comparisons, and calculate the normalized advantage value of each answer relative to the intra-group mean On this basis, the objective function of GRPO can be written as:

[0080] , in For the new and old strategies The probability ratio of tokens, is the advantage value obtained by scoring the responses within the group and normalizing them, with min Together, we limit the update range. is the hyperparameter of the shear operation, It is control The coefficient of the divergence constraint strength, is the reference strategy (set to LLaMA-32B).

[0081] Step S52: Iteratively update the knowledge base: Based on expert feedback, the latest industry literature, equipment maintenance manuals, and expert reports, the system regularly adds new failure cases and failure modes to the domain knowledge base. This identifies and integrates previously uncovered failure modes (such as new equipment failures or anomalies caused by process adjustments), continuously expanding and revising the knowledge base content to enhance LLM's understanding and generalization capabilities for specific industrial scenarios.

[0082] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art will understand that the present invention includes, but is not limited to, the contents described in the above specific embodiments. Any modifications that do not deviate from the functional and structural principles of the present invention are intended to be included within the scope of the claims.

Claims

1. An automated fault diagnosis method for industrial systems based on a large language model, characterized in that: Including steps: S1, constructing an industrial system operation data set and preprocessing the data in the industrial system operation data set; S2, builds semantic mapping relationships between industrial variables and converts industrial data into natural language descriptions; S3, builds a quantitative computing model based on a large language model, outputs abnormal content based on semantic fusion and comprehensive reasoning; S4, builds a knowledge reasoning model based on a large language model to diagnose industrial system faults; S5, optimizing the knowledge reasoning model based on the expert feedback mechanism, and adaptively updating the knowledge base and the knowledge reasoning model; The step S4 constructs a knowledge reasoning model based on a large language model to implement fault diagnosis of industrial systems, including: Step S41: Based on the retrieval-enhancement-generation technology, fault cases and diagnostic experiences that are highly relevant to the current abnormal state are retrieved in real time from the industrial system domain knowledge base constructed in step S14; Step S42: Based on the large language model, a comprehensive analysis is performed on the intrinsic correlation between the abnormal state and historical fault cases, a logical reasoning chain is generated, and a fault diagnosis conclusion is output; The step S41 includes: Word vector encoder using LLaMA Convert the data to be retrieved and each document into word vectors and : Use cosine similarity to calculate the similarity score between the data to be retrieved and each document ,in Represents a point operation, Represents the modulus of a vector: Select the one with the highest similarity Documents constitute a collection : Concatenate the exception information p with the k cases and use them together as the input q of the large language model: ； The step S5 optimizes the knowledge reasoning model based on the expert feedback mechanism, and adaptively updates the knowledge base and the knowledge reasoning model, including: Step S51, feedback optimization: establish a structured user feedback collection mechanism, record the operator's evaluation and scoring of the diagnosis conclusion, use the feedback data as input to perform reinforcement learning through the GRPO algorithm, and optimize the strategy of the knowledge reasoning model; Step S52, iterative update of knowledge base cases: based on expert feedback and the latest industry literature, equipment maintenance manuals and expert reports, new failure cases and failure modes are regularly added to the domain knowledge base.

2. The method for automated fault diagnosis of industrial systems based on a large language model according to claim 1, characterized in that: The step S1 comprises: Step S11: Obtain historical operating data of key equipment in the industrial system, external environmental factors at the corresponding time, physical parameters and load characteristics; construct an industrial system data set , ， in represents the measured data of the i-th feature, , represents the field of real numbers; Step S12: pre-process the data in the industrial system dataset X based on 3 The criterion for each column of features Conduct preliminary screening, including is the feature mean of the column, is the characteristic standard deviation of the column, and the observations that meet the range are recorded as : ; Based on the industrial business scope, the observations that meet the business scope are recorded as : , the rated operating range of the equipment is , The obtained data set Expressed as: ; Step S13, building an industrial system log library: the system collects raw log data from the production scheduling system, including process parameter change records, scheduling decisions, equipment status updates, and operator notes, and cleans, formats, and structures the log data; Step S14, building an industrial system domain knowledge base: sorting and collecting industrial system failure cases from public documents, industry reports, technical manuals, and historical failure logs, standardizing and storing the collected cases in a structured manner to form an industrial system domain knowledge base with multi-dimensional information such as time stamp, failure type, cause, diagnosis, and treatment plan.

3. The method for automated fault diagnosis of industrial systems based on a large language model according to claim 2, characterized in that: The step S2 comprises: Based on the multi-source monitoring variables obtained in step S11, an industrial variable-natural language description mapping table is constructed, the variables are matched with their standardized natural language descriptions, and a mapping relationship between the industrial variables and the standardized natural language is established.

4. The method for automated fault diagnosis of industrial systems based on a large language model according to claim 3, characterized in that: The step S3 constructs a quantitative calculation model based on a large language model to achieve abnormal content output, including: Step S31: construct a quantitative calculation model based on the large language model, generate a dynamic normal range for each variable, and preset reasonable boundaries of the variables; In step S32 , based on the dynamic normal range of the variable generated in step S31 , the real-time monitoring data is compared with the preset reasonable boundaries, and the specific values ​​of the parameters exceeding the normal range are output.

5. The method for automated fault diagnosis of industrial systems based on a large language model according to claim 4, characterized in that: A structured output mechanism is used to output the fault diagnosis conclusion including: Candidate fault set: The candidate fault set includes possible fault types and their confidence scores based on reasoning chain matching; Fault cause analysis: The fault cause analysis includes showing the logical relationship between abnormal parameters and revealing how they lead to specific faults; Recommended treatment plan: The recommended treatment plan provides targeted maintenance suggestions based on the case.

6. The method for automated fault diagnosis of industrial systems based on a large language model according to claim 5, characterized in that: The step S51 includes: The GRPO algorithm starts with the old strategy in an update. For the same exception information Sampling out Diagnosis results , compare the evaluation scores of the diagnostic conclusions within the group, and calculate the normalized advantage value of each answer relative to the group mean , on this basis, the objective function of GRPO is written as: in is the probability ratio of the new and old strategies on the tth token, is the advantage value obtained by scoring the responses within the group and normalizing them, with min Together, we limit the update range. is the hyperparameter of the shear operation, It is control The coefficient of the divergence constraint strength, It is a reference strategy.

7. An industrial system automated fault diagnosis system based on a large language model, characterized in that: Continuously collect key system equipment data in the industrial system as input, and output corresponding diagnosis results based on any industrial system automatic fault diagnosis method based on a large language model in claims 1-6.

Images