Offshore wind turbine fault diagnosis method based on thought chain fine tuning large model

Abstract

This invention discloses a fault diagnosis method for offshore wind turbine generators based on a large-scale model fine-tuned using a thought chain approach. Belonging to the field of wind turbine generator fault diagnosis, this method transforms collected SCADA and CMS data into higher-dimensional data features based on common signal characteristics in the field. It proposes constructing a fault diagnosis process with reasoning capabilities, adjusting the dataset to a thought chain dataset with a step-by-step reasoning process, and proportionally dividing it into training and test sets. The model is trained using the thought chain training dataset using a LoRa low-rank adaptive fine-tuning method. The trained model is then validated using a test set to obtain a high-accuracy fault diagnosis model. This invention significantly reduces labor costs. Furthermore, by combining the thought chain reasoning dataset, it ensures the reliability and efficiency of large-scale model reasoning, saving computational costs.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine fault diagnosis, specifically involving a fault diagnosis method for offshore wind turbines based on a large-scale model with fine-tuning of thought chains. Background Technology

[0002] Fault diagnosis of offshore wind turbines is an important research topic in the wind power industry. Due to the complex operating environment and diverse fault types of offshore wind turbines, traditional fault diagnosis methods often rely on feature engineering and rule models, which are difficult to use accurately and in real time. In recent years, with the development of neural network technology, intelligent diagnosis methods based on deep learning and large-scale data have gradually become a research hotspot.

[0003] However, existing intelligent diagnostic methods based on deep learning and large-scale data still have the following technical problems:

[0004] 1. Intelligent diagnostic methods based on deep learning and large-scale data rely on a large amount of high-quality data to train accurate diagnostic models.

[0005] 2. Existing intelligent diagnostic methods generally focus on feature extraction at the data level, lacking an understanding of semantics and context. Summary of the Invention

[0006] To address the shortcomings of existing technologies and improve the efficiency and accuracy of fault diagnosis for offshore wind turbine generators, this invention adopts the following technical solution:

[0007] The generator set fault diagnosis method based on the fine-tuning of a large model using the thought chain includes the following process:

[0008] The raw signals of the generator set are collected, and the time-series data composed of the raw signals are transformed into high-dimensional signal features that are strongly correlated with the occurrence of the fault, so as to characterize the equipment operating status when the fault occurs.

[0009] Based on the high-dimensional signal features and domain-related prior knowledge, a fault diagnosis prompt instruction set without a thought chain is constructed to communicate with the large model, obtain the large model's own reasoning process, adjust the large model's reasoning thought chain, extract, simplify and correct the original reasoning process of the large model, cut out the processes in the large model's thought chain that are useless to the diagnostic results, and re-plan the diagnostic process from feature data to fault results, transforming the fault diagnosis process into thought chain annotation to supplement the fault diagnosis prompt instruction set used for large model training and testing;

[0010] During the fine-tuning of the large model, the original parameter matrix of the large model is no longer updated. Only the introduced low-rank matrix is ​​optimized to fine-tune the pre-trained base model, resulting in the fine-tuned large model, which is then used for generator set fault diagnosis.

[0011] Furthermore, the original signal is preprocessed. For a data point, the number of other data points within the radius of the data point is obtained through the radius parameter and compared with the minimum number of points threshold. Density-based spatial clustering is then performed to divide the data points into core points, boundary points, and noise points in order to denoise the original signal.

[0012] The specific formula for the neighborhood of the radius parameter ε is as follows:

[0013]

[0014] in, ε represents the distance between points p and q, and ε is a predefined radius parameter that determines the neighborhood of point p.

[0015] The specific formula for determining the core point is as follows:

[0016]

[0017] in, MinPts represents the number of points in the ε-neighborhood of point p, and is a predefined minimum point threshold.

[0018] This method effectively removes outliers from signal data to improve data quality and prevent outlier data from affecting subsequent model training.

[0019] Furthermore, the high-dimensional signal features include average value, root mean square value, peak value, peak factor, impulse factor, and kurtosis.

[0020] The specific calculation formula is as follows:

[0021] average value: ;

[0022] Root mean square value: ;

[0023] Peak value: ;

[0024] Peak factor: ;

[0025] Pulse factor: ;

[0026] kurtosis: ;

[0027] in, This represents the input data, and N represents the number of data points.

[0028] By selecting highly correlated signal features, irrelevant features are avoided from excessively occupying the sequence length of the Prompt instruction set, thus affecting the training time and diagnostic performance of the model.

[0029] Furthermore, the thought chain of the large model includes the task instructions that the large model needs to complete, the reasoning process required for the large model to complete the task requirements, and the results obtained by the large model's reasoning. The reasoning process required for the large model to complete the task requirements includes the signal characteristics of fault data and health data in the domain knowledge of the large model, as well as the signal characteristics of the input of the large model's instruction set. By comparing the signal characteristics of fault data and health data in the domain knowledge with the differences of the input signal characteristics, the large model gradually judges the operating status of the equipment and finally obtains the equipment fault diagnosis results obtained by the large model's reasoning.

[0030] The specific format of the thought chain is as follows:

[0031]

[0032] in, The step-by-step reasoning process representing a large model, i.e., the thought chain. Represents the initial input of the large model. This represents the final output of the large model.

[0033] Thought chain for large models After simplification and modification, its definition formula is as follows:

[0034]

[0035] in, This represents the task instructions that the large model needs to complete. This represents the reasoning process required for a large model to complete the task requirements. This represents the result obtained from large-scale model reasoning.

[0036] For large model inference process Its specific definition formula is as follows:

[0037]

[0038] in, The signal characteristics of fault data within the larger model of knowledge; The signal characteristics of health data within the framework of large-scale model knowledge; The signal characteristics representing the input of the large model instruction set.

[0039] By eliminating processes in the large model's thought process that are useless to the diagnostic results and redesigning the diagnostic process from feature data to fault results, an efficient and accurate reasoning thought process for the large model can be obtained, effectively improving the model's training performance.

[0040] Furthermore, the instruction set includes user input and the output of a large model within the instruction set. The user input includes the role played by the large model, the task instructions that the large model needs to complete, and the contextual signal features input to the large model. The output of the large model within the instruction set includes the thought process of the large model and the output results of the large model. Through the contextual signal features, the high-dimensional signal features are fused into the instruction text. By assigning roles to the large model, the domain-related prior knowledge obtained by the large model's pre-training is used to gradually infer the device status. Through task instructions, the large model completes the fault diagnosis task and finally obtains the diagnosis result. This method abandons the problem of traditional fault diagnosis relying solely on signal processing without connecting with engineering practice and domain background context.

[0041]

[0042] Where x represents the instruction set; Represents user input within the instruction set. It is the role played by the designated large model. These are the task instructions that the large model needs to complete. These are the contextual signal features input into the large model; R represents the output of the large model in the instruction set.

[0043] Furthermore, the large model employs a low-rank adaptive fine-tuning method, introducing a first low-falling matrix and a second low-falling matrix. The output of the large model is the product of the input of the large model and the original parameter matrix, plus the product of the first low-falling matrix, the second low-falling matrix, and the input of the large model. The first low-falling matrix is ​​updated through Gaussian random initialization, and the second low-falling matrix is ​​initialized to 0.

[0044]

[0045] in, It is of scale The original model parameter matrix, , Let be two trainable parameter matrices, where For matrix sum matrix rank and much smaller .

[0046] The original parameter matrix is ​​no longer updated during model fine-tuning. Instead, it only optimizes the introduced low-rank matrix. sum matrix The final output data of the fine-tuned model is the sum of the two output data parts:

[0047]

[0048] in, and These are the model's output and input, respectively. The update matrix is ​​typically initialized and updated using Gaussian random initialization. Initialize to 0.

[0049] Furthermore, the method also includes: controlling the generation quality by controlling the temperature coefficient and sampling parameters, performing multiple rounds of queries on the fine-tuned large model, and obtaining the accuracy of the fine-tuned large model in judging wind turbine faults.

[0050] The LoRA method enables efficient updates to the original model without altering its parameter weights, significantly reducing the computational cost and training time required for training large models.

[0051] The method for fault diagnosis of offshore wind turbine generator sets based on the fine-tuning of a large model using the thinking chain method is used to fine-tune a large model of the offshore wind turbine generator set, and then the fine-tuned large model is used for fault diagnosis of the offshore wind turbine generator set.

[0052] A generator set fault diagnosis method based on fine-tuning a large model using the thought chain, applied to the edge:

[0053] The raw signals of the generator set are collected, and the time-series data composed of the raw signals are transformed into high-dimensional signal features that are strongly correlated with the occurrence of the fault, so as to characterize the equipment operating status when the fault occurs.

[0054] The high-dimensional signal features are sent to the server so that the server can construct a fault diagnosis instruction set without a thought chain based on the high-dimensional signal features and domain-related prior knowledge. This set can then communicate with the large model to obtain the large model's own reasoning process. This allows the server to adjust the large model's reasoning thought chain, remove processes in the large model's thought chain that are useless to the diagnosis results, and re-plan the diagnosis process. The fault diagnosis process is transformed into thought chain annotation to supplement the fault diagnosis instruction set used for training and testing the large model. During the fine-tuning of the large model, the original parameter matrix of the large model is no longer updated. Only the introduced low-rank matrix is ​​optimized to fine-tune the pre-trained base model, resulting in the fine-tuned large model.

[0055] The system collects the original signals of the generator set to be diagnosed and sends them to the server so that the server can perform generator set fault diagnosis using a trained large model and obtain the diagnosis results.

[0056] A generator set fault diagnosis method based on fine-tuning a large model using the mind chain, applied to the server side:

[0057] The time-series data of the generator set's original signals obtained from the edge are transformed into high-dimensional signal features that are strongly correlated with the occurrence of faults. Based on these high-dimensional signal features and domain-related prior knowledge, a fault diagnosis instruction set without a thought chain is constructed. This set interacts with the large model to obtain the large model's own reasoning process, thereby adjusting the large model's reasoning thought chain, eliminating processes in the large model's thought chain that are useless to the diagnostic results, and replanning the diagnostic process. The fault diagnosis process is transformed into thought chain annotation to supplement the fault diagnosis instruction set used for training and testing the large model. During the fine-tuning of the large model, the original parameter matrix of the large model is no longer updated; only the introduced low-rank matrix is ​​optimized to fine-tune the pre-trained base model, resulting in the fine-tuned large model.

[0058] The system acquires the original signals of the generator set to be diagnosed from the edge, performs fault diagnosis on the generator set using a trained large model, and obtains the diagnosis results.

[0059] The advantages and beneficial effects of this invention are as follows:

[0060] This invention addresses the complex operation and maintenance needs of offshore wind turbines, as well as the problems of high maintenance costs and complex fault types. It improves the traditional manual fault diagnosis method by transforming it into a data reasoning problem based on a large model, which greatly reduces labor costs. At the same time, by combining the thinking chain reasoning dataset, it ensures the reliability and efficiency of large model reasoning and saves computing costs. Attached Figure Description

[0061] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0062] Figure 2 This is a confusion matrix diagram of the test results of a large model fine-tuned on the training set in an embodiment of the present invention. Detailed Implementation

[0063] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0064] like Figure 1As shown, the fault diagnosis method for offshore wind turbines based on the fine-tuning of a large model using the thinking chain method employs the DBSCAN method and thresholding for data cleaning and noise reduction. Statistical analysis transforms the original time-frequency domain data into high-dimensional data feature values. Combined with expert prior knowledge, the fault diagnosis process is transformed into a thinking chain annotation supplemented into the fine-tuning Prompt instruction set. The large model is fine-tuned and trained using the LoRa method. Thus, considering the multi-sensor data dimensions of offshore wind turbines, and combining the thinking chain inference dataset with large model data alignment, the operating status of the offshore wind turbine is determined by utilizing the inference capability of the large model. Specifically, the method includes the following steps:

[0065] Step S1: For the data types acquired by the sensors, data cleaning methods are used to reduce noise and clean the data to improve the quality and accuracy of the sensor data. The specific data cleaning methods are the DBSCAN algorithm and the threshold method. The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) method controls the formation of clusters through two parameters: the radius parameter ε and the minimum number of points threshold MinPts, thereby dividing the data into three categories: core points, boundary points, and noise points.

[0066] The specific formula for the ε-neighborhood is as follows:

[0067]

[0068] in, ε represents the distance between points p and q, and ε is a predefined radius parameter that determines the neighborhood of point p.

[0069] The specific formula for determining the core point is as follows:

[0070]

[0071] in, MinPts represents the number of points in the ε-neighborhood of point p, and is a predefined minimum point threshold.

[0072] Step S2: Combining professional literature and work orders in this field, the cleaned time-series data is transformed into high-dimensional feature values ​​such as average value, peak value, and shape factor, which are strongly correlated with the occurrence of faults. These feature values ​​are used to accurately characterize the equipment operating status at the time of the fault. Specifically, the signal feature values ​​include six high-dimensional signal features: mean, root mean square value, peak value, peak factor, impulse factor, and kurtosis. The specific calculation formula is as follows:

[0073] average value: ;

[0074] Root mean square value: ;

[0075] Peak value: ;

[0076] Peak factor: ;

[0077] Pulse factor: ;

[0078] kurtosis: ;

[0079] in, This represents the input data, and N represents the number of data points.

[0080] Step S3: Based on the calculated high-dimensional signal feature values ​​and domain-related prior knowledge, construct a Prompt fault diagnosis instruction set without a thought chain, and communicate with the large model to obtain its own reasoning process; extract, simplify, and correct the original reasoning process of the large model, remove processes in the large model's thought chain that are useless to the diagnostic results, and re-plan the diagnostic process from feature data to fault results to obtain an accurate and efficient reasoning thought chain for the large model, thereby improving the large model's reasoning speed and diagnostic accuracy in this domain; the specific style of the thought chain is as follows:

[0081]

[0082] in, This represents the step-by-step reasoning process of a large model from data input to diagnostic results, i.e., the thought chain. Represents the initial input of the large model. This represents the final output of the large model.

[0083] Thought chain for large models After simplification and modification, its definition formula is as follows:

[0084]

[0085] in, This represents the task instructions that the large model needs to complete. This represents the reasoning process required for a large model to complete the task requirements. This represents the result obtained from large-scale model reasoning.

[0086] For large model inference process Its specific definition formula is as follows:

[0087]

[0088] in, The signal characteristics of fault data within the larger model of knowledge; The signal characteristics of health data within the framework of large-scale model knowledge; The signal characteristics representing the input of the large model instruction set are used by the large model to gradually determine the operating status of the equipment by comparing the differences between the fault and health signal characteristics in the domain knowledge and the input signals, and finally obtain the equipment fault diagnosis results.

[0089] Step S4: Based on the obtained high-dimensional signal features and the large model inference thought chain, construct a fault diagnosis prompt instruction set for model training and testing. The specific instruction set template is shareGPT, and its definition formula is as follows:

[0090]

[0091] Where x represents the instruction set; Represents user input within the instruction set. It is the role played by the designated large model. These are the task instructions that the large model needs to complete. These are the contextual signal features input into the large model; R represents the output of the large model in the instruction set. In this method, the problem of traditional fault diagnosis relying solely on signal processing without considering engineering practice and domain context is eliminated. Instead, the method integrates high-dimensional signal features into the instruction text and uses the domain knowledge of a large model to gradually infer the equipment status, ultimately obtaining the diagnostic result.

[0092] Step S5: Select an efficient parameter fine-tuning method and appropriate training parameters to fine-tune the pre-trained pedestal model; the large model fine-tuning method adopts the Low-Rank Adaptation (LoRA) method, the specific definition of which is as follows:

[0093]

[0094] in, It is of scale The original model parameter matrix, , Let be two trainable parameter matrices, where For matrix sum matrix rank and much smaller .

[0095] The original parameter matrix is ​​no longer updated during model fine-tuning. Instead, it only optimizes the introduced low-rank matrix. sum matrix The final output data of the fine-tuned model is the sum of the two output data parts:

[0096]

[0097] in, and These are the model's output and input, respectively. The update matrix is ​​typically initialized and updated using Gaussian random initialization. Initialize to 0.

[0098] Step S6: By controlling the generation quality through temperature coefficient and sampling parameters, perform multiple rounds of queries on the fine-tuned large model to obtain the accuracy of the fine-tuned large model in judging wind turbine faults.

[0099] To verify the effectiveness of the method of this invention, this embodiment uses a publicly available wind turbine gearbox vibration signal dataset, which includes two types of operating data: healthy and broken teeth. This dataset was recorded by sensors in four different directions under loads ranging from "0" to "90%". In this embodiment, the two types of operating data at 90% load were selected and used to create 100 training data points at a 1:1 ratio. The signal characteristics of each training data point were calculated from 5 seconds of operating data. The specific training set thought process is as follows:

[0100] “human”:

[0101] "You are a fault diagnosis expert in the wind power field. You can diagnose the operating status of equipment based on the given signal characteristics. Please determine the operating status of the equipment based on the provided gearbox vibration data.\nThe signal characteristics of the gearbox in four different directions\nThe average value is: [-0.04, 0.1, -0.02, 0.1]; the root mean square value is: [4.9, 5.2, 4.0, 4.9]; the peak value is: [12.8, 14.8, 12.9, 14.676]; the peak factor is: [3.3, 3.4, 4.1, 3.8]; the impulse factor is: [2.5, 2.8, 3.1, 2.9]; the kurtosis is: [0.03, -0.11, 0.76, 0.13]."

[0102] “GPT”:

[0103] “ <think> First, the user asked me to diagnose the operating status of the gearbox. I will determine the operating status of the equipment based on the signal characteristics provided by the user. Generally, when a broken tooth occurs, the peak value, peak factor, impulse factor, and other characteristic values ​​of the equipment are relatively large, and the kurtosis value is greater than 0; while when the equipment is operating normally, all characteristic values ​​are relatively small, and the kurtosis value is less than 0. Based on these characteristics, I can determine that the operating status of the equipment is a broken tooth.< / think> Broken teeth

[0104] Meanwhile, to verify the advantages of this method compared to traditional fault diagnosis methods, this invention selects a CNN convolutional neural network for comparative experiments and introduces metrics such as ACC and F1-meanssure to quantitatively represent the accuracy of the model. The specific calculation formula is as follows:

[0105]

[0106]

[0107]

[0108]

[0109] In this context, TP represents a correct fault prediction, TN represents a correct health prediction, FP represents an incorrect fault prediction, and FN represents an incorrect health prediction.

[0110] The confusion matrix after training on the training set is as follows: Figure 2 As shown in Table 1, the comparison results between the quantitative indicators of the large model and the convolutional neural network show that the large model can effectively understand and extract the hidden features in the vibration signal after a short period of thinking, and can effectively complete the task of diagnosing broken teeth in the gearbox of wind turbine units.

[0111] Table 1 Comparison of Indicators of Different Models

[0112] ACC F1-measure G-mean MCC Domain knowledge fine-tuning of large models 0.94 0.94 0.94 0.88 Convolutional Neural Networks 0.9138 0.9163 0.9133 0.8296

[0113] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A generator set fault diagnosis method based on a large model fine-tuning using a thought chain, characterized by: The raw signals of the generator set are collected, and the time-series data composed of the raw signals are transformed into high-dimensional signal features that are strongly correlated with the occurrence of the fault, so as to characterize the equipment operating status when the fault occurs. Based on the high-dimensional signal features and domain-related prior knowledge, a fault diagnosis instruction set without a thought chain is constructed to communicate with the large model and obtain the large model's own reasoning process. This allows for the adjustment of the large model's reasoning thought chain, the removal of processes in the large model's thought chain that are useless to the diagnostic results, and the replanning of the diagnostic process. The fault diagnosis process is transformed into thought chain annotation to supplement the fault diagnosis instruction set used for training and testing the large model. During the fine-tuning of the large model, the original parameter matrix of the large model is no longer updated. Only the introduced low-rank matrix is ​​optimized to fine-tune the pre-trained base model, resulting in the fine-tuned large model, which is then used for generator set fault diagnosis.

2. The generator set fault diagnosis method based on a large model fine-tuning using a thought chain as described in claim 1, characterized in that: The original signal is preprocessed. For a data point, the number of other data points within the radius of the data point is obtained through the radius parameter and compared with the minimum number of points threshold. Density-based spatial clustering is performed to divide the data points into core points, boundary points and noise points in order to denoise the original signal.

3. The generator set fault diagnosis method based on a large-scale model with fine-tuning of the thought chain as described in claim 1, characterized in that: The high-dimensional signal features include, but are not limited to, the average value, root mean square value, peak value, peak factor, impulse factor, and kurtosis of the input data.

4. The generator set fault diagnosis method based on a large model fine-tuning using a thought chain as described in claim 1, characterized in that: The thought process of the large model includes the task instructions that the large model needs to complete, the reasoning process required for the large model to complete the task requirements, and the results obtained by the large model's reasoning. The reasoning process required for the large model to complete the task requirements includes the signal characteristics of fault data and health data in the domain knowledge of the large model, as well as the signal characteristics of the input of the large model's instruction set. The large model judges the operating status of the equipment by comparing the signal characteristics of the fault data and health data in the domain knowledge with the differences of the input signal characteristics, and finally obtains the equipment fault diagnosis results obtained by the large model's reasoning.

5. The generator set fault diagnosis method based on a large-scale model fine-tuning using a thought chain as described in claim 4, characterized in that: The instruction set includes user input and the output of a large model within the instruction set. User input includes the role played by the large model, the task instructions that the large model needs to complete, and the contextual signal features input to the large model. The output of the large model in the instruction set includes the thought process of the large model and the output results of the large model. Through the contextual signal features, the high-dimensional signal features are fused into the instruction text. By assigning roles to the large model, the domain-related prior knowledge obtained by the large model's pre-training is used to gradually infer the device status. Through task instructions, the large model completes the fault diagnosis task and finally obtains the diagnosis result.

6. The generator set fault diagnosis method based on a large-scale model fine-tuning using a thought chain as described in claim 1, characterized in that: The large model employs a low-rank adaptive fine-tuning method, introducing a first low-falling matrix and a second low-falling matrix. The output of the large model is the product of the input of the large model and the original parameter matrix, plus the product of the first low-falling matrix, the second low-falling matrix, and the input of the large model. The first low-falling matrix is ​​updated through Gaussian random initialization, and the second low-falling matrix is ​​initialized to 0.

7. The generator set fault diagnosis method based on a large model fine-tuning using a thought chain as described in claim 1, characterized in that: The method further includes: controlling the generation quality by controlling the temperature coefficient and sampling parameters, performing multiple rounds of queries on the fine-tuned large model, and obtaining the accuracy of the fine-tuned large model in judging generator set faults.

8. A fault diagnosis method for offshore wind turbine generators based on a large-scale model with fine-tuning of thought chains, characterized by: The generator set fault diagnosis method based on the mental chain fine-tuning large model as described in any one of claims 1 to 7 is used to fine-tune the large model of the offshore wind turbine, and the fine-tuned large model is used for fault diagnosis of the offshore wind turbine.

9. A generator set fault diagnosis method based on a large model fine-tuning using a thought chain, applied to the edge end, characterized by: The raw signals of the generator set are collected, and the time-series data composed of the raw signals are transformed into high-dimensional signal features that are strongly correlated with the occurrence of the fault, so as to characterize the equipment operating status when the fault occurs. The high-dimensional feature signal is sent to the server so that the server can construct a fault diagnosis instruction set without a thought chain based on the high-dimensional signal features and domain-related prior knowledge. The server can then communicate with the large model to obtain the large model's own reasoning process, adjust the large model's reasoning thought chain, remove processes in the large model's thought chain that are useless to the diagnosis results, and re-plan the diagnosis process. The fault diagnosis process is transformed into thought chain annotation to supplement the fault diagnosis instruction set used for training and testing the large model. During the fine-tuning of the large model, the original parameter matrix of the large model is no longer updated. Only the introduced low-rank matrix is ​​optimized to fine-tune the pre-trained base model and obtain the fine-tuned large model. The system collects the original signals of the generator set to be diagnosed and sends them to the server so that the server can perform generator set fault diagnosis using a trained large model and obtain the diagnosis results.

10. A generator set fault diagnosis method based on a large model fine-tuning using a thought chain, applied to the server side, characterized by: The time-series data of the original generator set obtained from the edge end are transformed into high-dimensional signal features that are strongly correlated with the occurrence of faults. Based on the high-dimensional signal features and domain-related prior knowledge, a fault diagnosis instruction set without a thought chain is constructed. This set interacts with the large model to obtain the large model's own reasoning process. The reasoning thought chain of the large model is adjusted, processes that are useless to the diagnosis results are removed from the large model's thought chain, and the diagnosis process is re-planned. The fault diagnosis process is transformed into thought chain annotation to supplement the fault diagnosis instruction set used for training and testing the large model. During the fine-tuning of the large model, the original parameter matrix of the large model is no longer updated. Only the introduced low-rank matrix is ​​optimized to fine-tune the pre-trained base model and obtain the fine-tuned large model. The system acquires the original signals of the generator set to be diagnosed from the edge, performs fault diagnosis on the generator set using a trained large model, and obtains the diagnosis results.

Images