Blade fault diagnosis method and device based on large model, equipment, medium and program product

Abstract

The invention relates to the technical field of artificial intelligence data processing and the technical field of wind power generation, and provides a blade fault diagnosis method and device based on a large model, equipment, a medium and a program product. Generating vibration image data based on the vibration data; and performing fault diagnosis on the blade by using a first large model according to a preset prompt word and the vibration image data to obtain a fault diagnosis result, the preset prompt word prompting a diagnosis rule for performing fault diagnosis on the first large model. The problem that the blade fault diagnosis effect is difficult to improve is solved, the deep understanding and mining capability of the large model for data is utilized, the cue word is utilized to guide the large model to carry out fault diagnosis according to the rule given by the cue word, the fault diagnosis accuracy of the large model is improved, the method is more suitable for an intelligent wind power scene, and the fault diagnosis efficiency is improved. And digital and intelligent development of wind power operation and maintenance is promoted.

Description

Technical Field

[0001] This disclosure relates to the fields of artificial intelligence data processing technology and wind power generation technology. Specifically, it relates to a blade fault diagnosis method, device, equipment, medium, and program product based on a large model. Background Technology

[0002] As an important component of clean energy, wind power technology relies on the blades, a core component of wind turbine generators, whose health directly affects the generator's power generation efficiency, safety, and total life-cycle cost.

[0003] Current fault diagnosis technologies for wind turbine blades mainly include traditional methods based on vibration signals and diagnostic technologies based on general artificial intelligence (AI) models. Traditional methods typically use basic signal processing combined with statistical models or rule-based thresholds for fault diagnosis. However, their performance is highly dependent on carefully designed artificial features and static thresholds, and they cannot identify and diagnose faults based on an understanding of the fault data. While general AI models have a deep understanding of features, they also lack complex knowledge of specific domains, making it difficult to accurately identify and diagnose faults. Summary of the Invention

[0004] To address the problem that traditional methods and models lack the ability to understand fault data and general AI models lack complex knowledge of specialized fields, thus making it difficult to improve the blade fault diagnosis effect, this disclosure proposes a blade fault diagnosis method, device, equipment, medium, and program product based on a large model, which can solve or at least alleviate the above problems.

[0005] A first aspect of the embodiments of this disclosure provides a blade fault diagnosis method based on a large model. The blade fault diagnosis method includes: acquiring vibration data of a wind turbine blade; generating vibration image data based on the vibration data, wherein the vibration image data includes a time-domain graph and / or a spectrum graph; and using a first large model to perform fault diagnosis on the blade according to preset prompt words and the vibration image data to obtain a fault diagnosis result, wherein the preset prompt words prompt the first large model to perform fault diagnosis according to diagnostic rules.

[0006] Optionally, generating vibration image data based on the vibration data includes: filtering the vibration data to obtain filtered vibration data; generating the vibration image data based on the filtered vibration data, wherein the vibration data includes multiple time-series vibration data, and the data filtering includes: converting each time-series vibration data into a vibration spectrum; filtering the multiple time-series vibration data according to the amplitude variation trend of each vibration spectrum to obtain filtered time-series vibration data.

[0007] Optionally, the step of filtering the multiple time-series vibration data according to the amplitude variation trend of each vibration spectrum to obtain filtered time-series vibration data includes: performing envelope fitting on the amplitude of each vibration spectrum to obtain amplitude boundaries; for each vibration spectrum, determining the data variation trend of the amplitude boundary of the vibration spectrum within each moving window by moving the frequency window multiple times on the vibration spectrum; determining whether the data quality of the time-series vibration data is abnormal based on the consistency of the data variation trend within the multiple moving windows; and obtaining the filtered time-series vibration data by removing data with abnormal data quality from the multiple time-series vibration data.

[0008] Optionally, determining whether the data quality of the time-series vibration data is abnormal based on the consistency of the data change trends within the multiple moving windows includes: determining that the data quality of the time-series vibration data is abnormal in response to the fact that the data change trends of the target window in the multiple moving windows are all downward and the effective data values ​​within the preset frequency range are greater than preset values, wherein the target window is a window with an upward or downward data change trend.

[0009] Optionally, the blade fault diagnosis method further includes: responding to the fault diagnosis result indicating the existence of a fault of the target fault type, using a preset mechanism model based on the vibration data to perform fault diagnosis on the blade and obtain a verification diagnosis result, wherein the mechanism model is a model for diagnosing the target fault type using the operating principle of a wind turbine generator; responding to the verification diagnosis result indicating the existence of a fault of the target fault type, outputting the fault diagnosis result; responding to the verification diagnosis result indicating the absence of a fault of the target fault type, updating the preset prompt words based on the vibration data and the verification diagnosis result.

[0010] Optionally, the preset prompt words are determined by: acquiring a first dataset, wherein the first dataset includes multiple first data samples and a first sample label corresponding to each first data sample, the first sample label indicating the fault type of the corresponding first data sample; using a second large model, based on the first dataset, extracting fault features corresponding to each fault type and generating initial prompt words; using the first large model, optimizing the initial prompt words to determine the preset prompt words.

[0011] Optionally, optimizing the initial prompt words using the first large model to determine the preset prompt words includes: acquiring a second dataset, wherein the second dataset includes multiple second data samples and a second sample label corresponding to each second data sample, the second sample label indicating the fault type of the corresponding second data sample; performing at least one round of prompt word optimization on the initial prompt words using the first large model and the second dataset to determine the preset prompt words, wherein the prompt word optimization includes: using the first large model, performing fault diagnosis on the blade based on the prompt words to be optimized in this round and the multiple second data samples to obtain a test diagnosis result; obtaining a comparison result by comparing the test diagnosis result with the second sample label; determining the preset prompt words based on the prompt words to be optimized in this round in response to the comparison result satisfying the preset optimization conditions; and optimizing the prompt words to be optimized in this round in response to the comparison result not satisfying the preset optimization conditions, and using the optimized prompt words as the prompt words to be optimized in the next round.

[0012] Optionally, the prompt words to be optimized in this round are optimized in the following way: identify the target data samples that were missed or misjudged in the plurality of second data samples in this round of fault diagnosis; update the first dataset using the target data samples and the corresponding second sample labels to obtain the updated first dataset; and optimize the current prompt words based on the updated first dataset using the second large model to obtain the optimized prompt words for this round.

[0013] Optionally, the step of using the second large model to optimize the current prompt word based on the updated first dataset to obtain the optimized prompt word for this round includes: generating guidance information based on the target data sample and the current prompt word; and using the second large model to optimize the current prompt word based on the updated first dataset and the guidance information to obtain the optimized prompt word for this round, wherein the guidance information indicates the optimization direction for the second large model to optimize the current prompt word.

[0014] Optionally, generating guidance information based on the target data sample and the current prompt word includes: in response to a preset defect in the description of the fault type of the target data sample in the current prompt word, generating guidance information corresponding to the preset defect, wherein the preset defect includes at least one of the following: ambiguous feature description; overlapping diagnostic rules for different fault types; lack of diagnostic rules for fault edges, and the guidance information includes at least one of the following: extracting quantitative features; supplementing differential features; adding diagnostic rules for fault edges.

[0015] Optionally, the comparison results include: a precision metric for each fault type; and / or, a recall metric for each fault type, wherein the precision metric is determined based on the number of samples correctly diagnosed as the fault type and the total number of samples diagnosed as the fault type in the plurality of second data samples; and the recall metric is determined based on the number of samples correctly diagnosed as the fault type and the total number of samples of the fault type in the plurality of second data samples.

[0016] A second aspect of the embodiments of this disclosure provides a blade fault diagnosis device based on a large model, the blade fault diagnosis device comprising: an acquisition unit configured to acquire vibration data of a wind turbine blade; a generation unit configured to generate vibration image data based on the vibration data, wherein the vibration image data includes a time-domain plot and / or a spectrum plot; and a diagnosis unit configured to use a first large model to perform fault diagnosis on the blade according to preset prompt words and the vibration image data, and obtain a fault diagnosis result, wherein the preset prompt words prompt the first large model to perform fault diagnosis according to the diagnostic rules.

[0017] A third aspect of the embodiments of this disclosure provides an electronic device, the electronic device comprising: a processor; and a memory for storing processor-executable instructions, wherein, when executed by the processor, the processor causes the processor to perform a blade fault diagnosis method based on a large model according to the present disclosure.

[0018] A fourth aspect of the embodiments of this disclosure provides a wind turbine generator set, the wind turbine generator set including electronic equipment according to the present disclosure, or the wind turbine generator set being communicatively connected to electronic equipment according to the present disclosure.

[0019] A fifth aspect of the embodiments of this disclosure provides a computer-readable storage medium that, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform a blade fault diagnosis method based on a large model according to this disclosure.

[0020] A sixth aspect of the embodiments of this disclosure provides a computer program product including computer-executable instructions, optionally which, when executed by at least one processor, implement the blade fault diagnosis method based on a large model according to this disclosure.

[0021] According to the embodiments of the present disclosure, the blade fault diagnosis method and apparatus based on a large model, wind turbine generator set, equipment, medium, and program product can use a large model to perform fault diagnosis on the blades of a wind turbine generator set based on preset prompt words and vibration image data generated based on blade vibration data, and obtain fault diagnosis results. The prompt words can prompt the large model to perform fault diagnosis according to the diagnostic rules. In this way, on the one hand, the deep understanding and mining capabilities of the large model can be used to improve the intelligence of fault diagnosis; on the other hand, the preset prompt words can be used to guide the large model to perform blade fault diagnosis according to the diagnostic rules given by the prompt words, thereby improving the accuracy of the large model in fault diagnosis.

[0022] Furthermore, in the blade fault diagnosis scheme based on a large model according to the embodiments of this disclosure, using a large model to diagnose blade vibration data can overcome the limitations of traditional methods in terms of diagnostic rules and thresholds. By leveraging the data sensitivity and feature capture capabilities of the large model, it is possible to deeply understand the blade state information conveyed in the blade vibration data, making it more suitable for smart wind power scenarios and conducive to promoting the digital and intelligent development of wind power operation and maintenance. Attached Figure Description

[0023] Figure 1 This is a schematic flowchart illustrating a large-model-based blade fault diagnosis method according to an exemplary embodiment of the present disclosure.

[0024] Figure 2 This is a time-domain diagram showing normal vibration data of a blade in the oscillation direction according to an exemplary embodiment of the present disclosure.

[0025] Figure 3 This is a time-domain plot showing normal vibration data of a blade in the flapping direction according to an exemplary embodiment of the present disclosure.

[0026] Figure 4 This is a frequency domain diagram showing normal vibration data of a blade in the oscillation direction according to an exemplary embodiment of the present disclosure.

[0027] Figure 5 This is a frequency domain diagram showing normal vibration data of a blade in the flapping direction according to an exemplary embodiment of the present disclosure.

[0028] Figure 6 This is a time-domain plot of vibration data showing abnormal data quality in the oscillation direction of a blade according to an exemplary embodiment of the present disclosure.

[0029] Figure 7 This is a time-domain plot of vibration data showing abnormal data quality in the flapping direction of a blade according to an exemplary embodiment of the present disclosure.

[0030] Figure 8 This is a frequency domain diagram showing vibration data with abnormal data quality in the oscillation direction of a blade according to an exemplary embodiment of the present disclosure.

[0031] Figure 9 This is a frequency domain diagram showing vibration data with abnormal data quality in the flapping direction of a blade according to an exemplary embodiment of the present disclosure.

[0032] Figure 10 This is a schematic flowchart illustrating filtered temporal vibration data obtained in a large-model-based blade fault diagnosis method according to an exemplary embodiment of the present disclosure.

[0033] Figure 11 This is a schematic diagram illustrating the vibration spectrum and amplitude boundaries according to an exemplary embodiment of the present disclosure.

[0034] Figure 12 This is a time-domain plot showing fault vibration data of a blade in the oscillation direction according to an exemplary embodiment of the present disclosure.

[0035] Figure 13 This is a time-domain plot showing fault vibration data of a blade in the flapping direction according to an exemplary embodiment of the present disclosure.

[0036] Figure 14 This is a frequency domain diagram showing fault vibration data of a blade in the oscillation direction according to an exemplary embodiment of the present disclosure.

[0037] Figure 15 This is a frequency domain diagram showing fault vibration data of a blade in the flapping direction according to an exemplary embodiment of the present disclosure.

[0038] Figure 16 This is a schematic flowchart illustrating the determination of prompt words in a large-model-based blade fault diagnosis method according to an exemplary embodiment of the present disclosure.

[0039] Figure 17 This is a schematic flowchart illustrating prompt word optimization in a large-model-based blade fault diagnosis method according to an exemplary embodiment of the present disclosure.

[0040] Figure 18 This is a schematic flowchart illustrating the optimization of prompt words by updating the dataset in a large-model-based blade fault diagnosis method according to an exemplary embodiment of the present disclosure.

[0041] Figure 19 This is a flowchart illustrating an example of prompt word optimization according to an exemplary embodiment of the present disclosure.

[0042] Figure 20 This is a flowchart illustrating the overall framework of a large-model-based blade fault diagnosis method according to exemplary embodiments of the present disclosure. Detailed Implementation

[0043] The following detailed embodiments are provided to aid the reader in gaining a comprehensive understanding of the methods, apparatus, and / or systems described herein. However, various changes, modifications, and equivalents of the methods, apparatus, and / or systems described herein will become apparent upon understanding this disclosure. For example, the order of operations described herein is merely illustrative and is not limited to those orders set forth herein, but may be changed as will become clear upon understanding this disclosure, except for operations that must occur in a specific order. Furthermore, for clarity and conciseness, descriptions of features known in the art may be omitted.

[0044] The features described herein may be implemented in different forms and should not be construed as limited to the examples described herein. Rather, the examples described herein are provided only to illustrate some of the many feasible ways of implementing the methods, apparatus, and / or systems described herein, which will become clear upon understanding the disclosure of this application.

[0045] As used herein, the term “and / or” includes any one of the associated listed items and any combination of any two or more.

[0046] Although terms such as “first,” “second,” and “third” may be used herein to describe various components, assemblies, regions, layers, or parts, these components, assemblies, regions, layers, or parts should not be limited by these terms. Rather, these terms are used only to distinguish one component, assembly, region, layer, or part from another. Thus, without departing from the teaching of the examples described herein, the first component, first assembly, first region, first layer, or first part referred to as the first component, first assembly, first region, first layer, or first part may also be referred to as the second component, second assembly, second region, second layer, or second part.

[0047] In the specification, when an element (such as a layer, region, or substrate) is described as being "on" another element, "connected to," or "bonded to" another element, the element may be directly "on" another element, directly "connected to," or "bonded to" the other element, or one or more other elements may be present in between. Conversely, when an element is described as being "directly on" another element, "directly connected to," or "directly bonded to" another element, no other elements may be present in between.

[0048] The terminology used herein is for the purpose of describing various examples only and is not intended to limit disclosure. Unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. The terms “comprising,” “including,” and “having” indicate the presence of the described features, quantities, operations, components, elements, and / or combinations thereof, but do not preclude the presence or addition of one or more other features, quantities, operations, components, elements, and / or combinations thereof.

[0049] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains upon understanding this disclosure. Unless expressly defined herein, terms (such as those defined in a general dictionary) shall be interpreted as having a meaning consistent with their meaning in the context of the relevant field and in this disclosure, and shall not be interpreted in an idealized or overly formalistic manner.

[0050] Furthermore, in the description of the examples, detailed descriptions of well-known related structures or functions will be omitted when it is believed that such detailed descriptions would lead to a vague interpretation of this disclosure.

[0051] As mentioned earlier, in blade fault diagnosis solutions, traditional methods and models lack the ability to understand fault data, while general AI models lack complex knowledge of specialized fields, which makes it difficult to improve the effectiveness of blade fault diagnosis.

[0052] Specifically, some fault diagnosis methods based on vibration signals mainly rely on signal processing (such as wavelet packet decomposition and Fourier transform), feature extraction (such as energy entropy, coherence coefficient, and coupled causality measure), and machine learning models (such as XGBoost and vector regression models) for diagnosis. These methods can identify mid-to-late stage faults in blades to some extent or filter out some quality problems through signal jump detection. However, they generally suffer from low diagnostic rate (or accuracy) and recall.

[0053] For example, in addition to signal fluctuations, sensor parameter errors, loose cables, environmental noise, and especially interference vibrations of the unit under different operating conditions (such as start-up, shutdown, low speed, and yaw) can all affect the stability of feature extraction and fault discrimination thresholds, making the diagnostic results highly sensitive to data quality.

[0054] For example, conventional models are poorly adaptable to complex operating conditions and different wind turbine models. In particular, for weak, early structural damage (such as microcracks), the vibration characteristics caused by these damages are subtle and nonlinear, making it difficult for traditional features and thresholds to effectively capture them, resulting in insufficient early warning capabilities.

[0055] Furthermore, some purely data-driven "black box" models fail to systematically encode the physical and operational mechanisms of blade failures into the model, affecting the credibility of diagnostic conclusions and their engineering interpretability.

[0056] While some fault diagnosis methods that employ general AI models demonstrate strong potential in learning general features, they still face challenges when directly applied to the vibration diagnosis of wind turbine blades with high signal-to-noise ratios and complex mechanisms.

[0057] For example, the model has a biased understanding of industrial characteristics, making it difficult to accurately understand the physical and operational mechanisms inherent in vibration signals (such as the rise of frequency harmonic energy due to crack faults, and the relationship between imbalance and rotational speed). This results in insufficient ability to identify actual fault characteristics and makes it easy to confuse with data quality issues.

[0058] For example, high-precision large-scale model inference relies on massive computing resources, which are difficult to deploy directly and in real time at the edge of wind turbines or in cost-sensitive scenarios, resulting in high computing resources and deployment costs.

[0059] In summary, in the identification and diagnosis of blade vibration in wind turbine generators, there is a disconnect between "traditional models that are mechanistic but have limited capabilities" and "AI large models that are powerful but lack expertise in complex domain knowledge." A complete and efficient end-to-end collaborative diagnosis solution has not yet been formed, encompassing source data governance, mechanism-guided multimodal large models, and resource-based collaborative optimization.

[0060] In view of this, the present disclosure provides a blade fault diagnosis method based on a large model, a blade fault diagnosis device based on a large model, a wind turbine generator, an electronic device, a computer-readable storage medium, and a computer program product, which can solve or at least alleviate at least one of the above problems.

[0061] According to an exemplary embodiment of the present disclosure, a blade fault diagnosis method based on a large model is provided. The method can be executed by an electronic device, such as a wind turbine blade vibration data, generating vibration image data based on the vibration data, wherein the vibration image data includes a time domain diagram and / or a spectrum diagram, and using a first large model, performing fault diagnosis on the blade according to preset prompt words and the vibration image data to obtain a fault diagnosis result, wherein the preset prompt words prompt the first large model to perform fault diagnosis according to the diagnostic rules.

[0062] In this way, on the one hand, the deep understanding and mining capabilities of large models can be used to improve the intelligence of fault diagnosis; on the other hand, preset prompts can be used to guide large models to perform blade fault diagnosis according to the diagnostic rules given by the prompts, thereby improving the accuracy of fault diagnosis by large models.

[0063] The embodiments of the blade fault diagnosis method based on a large model according to the present disclosure will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, this blade fault diagnosis method based on a large model may include the following steps: In step S110, vibration data of the wind turbine blades can be obtained.

[0064] As an example, blade vibration data can be acquired by a vibration sensor installed on the blade. For instance, the vibration sensor can acquire vibration signals in the blade's flapping and oscillating directions in real time. However, the embodiments of this disclosure are not limited to this, and vibration data can also be obtained through other direct measurement or indirect calculation methods.

[0065] In addition, as an example, operating data of wind turbine generators, such as SCADA (Supervisory Control and Data Acquisition) data, can be obtained. This operating data can be used to obtain real-time operating condition data, including but not limited to active power and yaw speed, so that vibration data can be screened for operating condition validity based on SCADA data.

[0066] In step S120, vibration image data can be generated based on the vibration data.

[0067] Here, vibration image data may include time-domain plots and / or spectrograms.

[0068] In one example, vibration data may include time-series vibration data (e.g., vibration waveform data), which can be converted into time-frequency images, such as time-domain plots and / or spectrum plots. The time-domain plot may contain the characteristics of the blade vibration in the time domain, and the spectrum plot may contain the characteristics of the blade vibration in the frequency domain. In subsequent steps, fault diagnosis can be performed based on the time-domain features in the time-domain plot and / or the frequency-domain features in the spectrum plot.

[0069] In another example, vibration data can be filtered first before generating vibration image data.

[0070] Specifically, step S120 may include: filtering the vibration data to obtain filtered vibration data; and generating vibration image data based on the filtered vibration data.

[0071] As an example, data filtering may include time-domain filtering and / or frequency-domain filtering.

[0072] Time-domain filtering can include, but is not limited to: removing shutdown / standby data, removing transition data, and removing data with incomplete time series.

[0073] Specifically, as mentioned above, operational data such as SCADA data of wind turbine generators can be acquired, and real-time operating condition data (such as active power and yaw speed) can be synchronously obtained from the operational data to perform time-domain operating condition filtering on vibration data based on the operational data.

[0074] For example, data with active power greater than a preset power (e.g., 50 kW) and yaw speed greater than 0 can be selected, while vibration signals from shutdown or standby states and yaw periods can be removed. After screening, vibration data sequences under stable power generation conditions are retained and entered into subsequent processing. In this way, through multi-source data synchronous acquisition and operating condition screening, the problem of model performance degradation caused by reduced signal-to-noise ratio of vibration data due to operating condition interference in related technologies can be effectively solved.

[0075] In addition, temporal filtering can also include: preliminary filtering of data quality issues through lightweight conventional models deployed at the edge, such as combining sliding window statistics (such as mean and standard deviation) to identify and remove drastic signal jumps and zeroing points caused by sensor failures and transient interference.

[0076] Frequency domain filtering can include, for example, conventional frequency domain data filtering. Furthermore, as an example, embodiments of this disclosure also propose a data quality depth filtering method based on a spectral skid slope algorithm.

[0077] As an example, vibration data may include multiple time-series vibration data. The data filtering described above may include: converting each time-series vibration data into a vibration spectrum; filtering the multiple time-series vibration data according to the changing trend of the amplitude of each vibration spectrum to obtain filtered time-series vibration data.

[0078] Specifically, when the blade sensor initially fails, the spectrum will exhibit a skid-slope-like data pattern. The abnormal vibration data will have a large number of random spikes and sudden peaks in the time-domain waveforms of the blade edge and blade flap channels. Normal periodic vibrations will be submerged by noise, resulting in chaotic waveforms, extremely low signal-to-noise ratio, and signals that deviate from smooth vibration patterns. Abnormal pulses will occur frequently, and the integrity of the time-domain data will be poor, making it impossible to clearly extract effective vibration features.

[0079] For example, the spectrum exhibits extremely high and diffuse energy in the low-frequency range (e.g., 0-200Hz), with no obvious dominant frequency peaks. In the mid-to-high frequency range (e.g., 200-1200Hz), persistent noise energy remains, the overall spectral floor is raised, the spectrum of each channel is irregularly distributed, there are no clear resonant characteristic peaks, noise is diffused across the entire frequency range, and some channels show local abnormal peaks, reflecting severe noise contamination and abnormal data quality. Normal data in the time domain exhibits similar characteristics. Figure 2 and Figure 3As shown, normal data behaves in the frequency domain (e.g., 0~80Hz) for example... Figure 4 and Figure 5 As shown, vibration data with abnormal data quality exhibits the following characteristics in the time domain: Figure 6 and Figure 7 As shown, vibration data with abnormal data quality exhibits the following characteristics in the frequency domain (e.g., 0~80Hz): Figure 8 and Figure 9 As shown, in Figures 2 to 9 In the diagram, blade oscillation 1, blade oscillation 2, and blade oscillation 3 represent the data of the three blades in the oscillation direction, respectively, while blade flapping 1, blade flapping 2, and blade flapping 3 represent the data of the three blades in the flapping direction, respectively.

[0080] In the embodiments of this disclosure, multiple time-series vibration data can be converted into corresponding vibration spectra respectively. Based on the changing trend of the amplitude of each vibration spectrum, vibration data with abnormal data quality can be identified and removed from the multiple time-series vibration data.

[0081] As an example, such as Figure 10 As shown, the steps described above for filtering multiple time-series vibration data based on the amplitude variation trend of each vibration spectrum to obtain filtered time-series vibration data may include the following steps: In step S1010, the amplitude of each vibration spectrum can be envelope-fitted to obtain the amplitude boundary.

[0082] As an example, given the vibration spectrum corresponding to each time series vibration data, the upper boundary of the vibration spectrum amplitude can be fitted using existing envelope fitting methods.

[0083] For example, peak suppression processing can be applied to the vibration spectrum, which refers to removing extreme peaks from the spectral waveform. For the original vibration spectrum or the peak-suppressed vibration spectrum, methods such as quantile-based filtering and linear interpolation can be used to generate a smooth envelope that fits the upper boundary of the vibration spectrum amplitude. For example, ... Figure 11 As shown, the blue line represents the original frequency domain data (or vibration spectrum), and the red curve represents the upper boundary curve of the fit (or the amplitude boundary of the fit).

[0084] In this step, envelope fitting can be performed for each vibration spectrum to obtain the amplitude boundary of each vibration spectrum.

[0085] In step S1020, for each vibration spectrum, the amplitude boundary of the vibration spectrum can be determined by moving the frequency window multiple times within each moving window to determine the data change trend of the vibration spectrum.

[0086] Here, the window size of the frequency window and the step size for each frequency window movement can be preset according to actual needs.

[0087] In this step, multiple windows can be selected on each vibration spectrum by sliding or moving the frequency window multiple times, and the data change trend of the amplitude boundary within each window can be determined. Here, the data change trend can include an upward trend, a downward trend, and no trend. No trend can mean, for example, that the data does not change significantly within the window, or is always within a predetermined amplitude range, or that the maximum difference of the amplitude boundary values ​​within the window is less than a preset difference.

[0088] Here, the method for determining the data change trend within each window at the amplitude boundary can be, for example, existing data trend determination methods, such as determining the data change trend by determining the slope of the curve within the window. However, the embodiments of this disclosure are not limited to this, and other methods can also be used.

[0089] In step S1030, the data quality of the time-series vibration data can be determined as abnormal based on the consistency of the data change trend within the multiple moving windows.

[0090] Here, the consistency of data change trends can refer to whether the data change trends are all the same.

[0091] As an example, step S1030 may include: in response to the fact that the data change trend of the target window in multiple moves is all downward and the effective value of the data within the preset frequency range is greater than a preset value, determining that the data quality of the time-series vibration data is abnormal. Here, the effective value of the data may be, for example, a representative value or typical value of the data, which can be determined according to actual needs, such as the square root or root mean square value of the data.

[0092] Here, the target window is the window whose data trend is rising or falling. For example, the target window can be selected from multiple moving windows, showing a non-trending data trend, and then a trend consistency judgment can be made based on the target window.

[0093] As an example, before performing the step of determining that the data quality of the time-series vibration data is abnormal when the data change trend of the target window in the multiple-movement window is all downward and the effective value of the data within the preset frequency range is greater than the preset value, the step S1030 may further include: determining the proportion of the first window in the multiple-movement window where the data change trend is trendless; and in response to the proportion being less than a preset proportion threshold, determining the windows other than the first window in the multiple-movement window as the target window.

[0094] The above-mentioned preset ratio threshold can be set according to actual needs, for example, it can be 35%.

[0095] Furthermore, as an example, before determining the target window, the first and last windows can be discarded from the windows that have been moved multiple times, and the proportion of the first window in the remaining windows can be determined.

[0096] As an example, the proportion of the first window can be the ratio of the number of first windows with no trend in data change to the total number of windows that have been moved multiple times; or, in the example of discarding the first and last windows, the proportion of the first window can be the ratio of the number of first windows with no trend in data change to the total number of remaining windows after discarding the first and last windows.

[0097] In the embodiments of this disclosure, considering that the data spectrum will exhibit a skid-slope pattern when the blade sensor initially fails, in step S1030, it can be checked whether the data change trend of the target window is entirely downward, and it can also be checked whether the effective data value within a preset frequency range is greater than a preset value. Here, the preset frequency range can be, for example, 10~600Hz, and the preset value can be, for example, 0.3. If the data change trend of the target window is entirely downward and the effective data value within the preset frequency range is greater than the preset value, it can be considered that the vibration data corresponding to the amplitude boundary has abnormal data quality.

[0098] The above judgment process can be performed for each amplitude boundary (or for each time series vibration data) to find data with abnormal data quality.

[0099] In step S1040, filtered time-series vibration data can be obtained by removing data with abnormal data quality from multiple time-series vibration data.

[0100] In this step, data with abnormal data quality can be removed from multiple time-series vibration data to obtain filtered time-series vibration data.

[0101] The above describes data quality deep filtering based on the spectrum ski slope algorithm. According to embodiments of this disclosure, frequency domain filtering may alternatively or additionally include anomalous noise filtering.

[0102] Specifically, abnormal noise filtering refers to using a bandpass filter to filter out broadband environmental noise that is unrelated to the blade's rotational frequency and its harmonics, based on the current impeller speed information.

[0103] By performing abnormal noise filtering, high-quality, clean blade vibration signals can be output (e.g., as mentioned above). Figures 2 to 5 The time-domain and spectrum diagrams of normal vibration data are shown.

[0104] In addition, if an unrepairable serious quality problem is detected during the above data filtering, a local data anomaly alarm can be triggered directly, pausing or ending the subsequent fault diagnosis process.

[0105] Return to reference Figure 1 In step S130, the first large model can be used to perform fault diagnosis on the blade based on preset prompt words and vibration image data to obtain fault diagnosis results.

[0106] Here, the aforementioned preset prompts can guide the first major model in performing fault diagnosis rules, and the prompts can be natural language text. Examples of these preset prompts and their determination methods will be described in detail below.

[0107] As an example, in step S130, vibration image data generated based on the high-quality vibration signal after the above data filtering and preset prompt words, the two core multimodal data, can be input into the first large model.

[0108] Here, the first large model can be, for example, a multimodal large model with image and natural language understanding capabilities. The first large model can be, for example, the open-source large model "Qwen3-VL-30B-A3B-Instruct", but the embodiments of this disclosure are not limited to this, and other multimodal large models can also be selected for testing and replacement.

[0109] In step S130, the first major model, guided by prompts, focuses on the features and patterns in the vibration image data and outputs fault diagnosis results for the blade, such as determining whether a fault exists and classifying the fault. As an example, the first major model can use a cross-modal attention mechanism to deeply fuse visual features with the mechanistic logic of text descriptions, ultimately outputting structured fault diagnosis results, which can be stored in a database.

[0110] As an example, the fault diagnosis results output by the first major model can strictly follow a predefined business prompt word template, which may include, but is not limited to: file name (file), diagnostic analysis (analysis), diagnostic result (answer), fault type (fault_type), and confidence level (confidence). Furthermore, the fault diagnosis results may, but are not limited to, be in JSON format. Here, the diagnostic analysis may contain the basis for the major model's current diagnostic results and fault type, and the confidence level characterizes the reliability of the fault diagnosis results.

[0111] For example, the time-domain plot and spectrum plot generated from the vibration data collected when a blade fails are as follows: Figures 12 to 15 As shown, where, Figure 12 and Figure 13Examples of time-domain plots of fault vibration data for the blade in the flaring and flapping directions are shown respectively. Figure 14 and Figure 15 Examples of frequency domain plots of fault vibration data for the blade in the flaring and flapping directions are shown respectively. The above... Figures 12 to 15 The following is an example of the fault diagnosis results output after inputting into the first large model: {"file":"Time-domain waveform.jpg,Spectrum plot.jpg", "analysis": "Time domain: Blade oscillations 1 / 2 / 3 exhibit disordered vibration coordination, irregular phase differences, and chaotic curve intersections; Spectrum: Abnormally high amplitude peaks appear in local frequency bands, and the spectral characteristics of each blade differ significantly (e.g., blade oscillation 2 has a large difference in spectrum from other blades, and blade flapping 2 has more clutter)". "answer":"1", "fault_type":"Structural looseness","confidence":"0.9"}.

[0112] The following will describe in detail examples of the preset prompt words and their determination methods.

[0113] As an example, prompt words can adopt a modular design that includes fixed and optimized parts to ensure that format constraints remain unchanged and fault determination rules can be iteratively optimized.

[0114] As an example, fixed parts of cue words can be used to define the output format, task objectives, and core requirements of a large model, preventing the output from deviating from the task scope. Below is an example of a fixed part of a cue word: text The analysis images are vibration waveform time-series diagrams and spectrum diagrams of three blades of a wind turbine. The task is to complete "normal / abnormal determination," "fault type classification," and "confidence level," and return the results in the specified JSON format. No extra text, punctuation, code blocks, or formatting. Image names should directly use the original jpg / png endings, and the analysis and description should only extract the core judgment criteria without digressing.

[0115] Strictly return this format: {"file":"Original image name","analysis":"Core judgment criteria (time domain + spectral characteristics)","answer":"0 / 1","fault_type":"Specific fault type / normal","confidence":"Confidence level"}.

[0116] In addition, the fixed part may also include key constraints, such as those used to emphasize "no additional text" and / or "only extract the core evidence," to avoid redundant content in the model output. Furthermore, key constraints may also include, for example, that field names and value ranges (answer is 0 / 1) in the JSON format must be strictly fixed.

[0117] As an example, the optimized portion of the prompt words can be adjusted during the prompt word optimization process. This optimized portion can be used to define the feature mapping between fault determination criteria and fault types.

[0118] The prompt words are obtained by splicing and integrating the fixed and optimized parts. For example, an initial version of the prompt words can be obtained first in the manner described above, and then optimized using the prompt word optimization mechanism or closed-loop iteration mechanism described below to obtain the final prompt words.

[0119] As an example, such as Figure 16 As shown, preset prompt words can be determined in the following ways: In step S1610, the first dataset can be obtained.

[0120] Here, the first dataset (hereinafter also referred to as the case set) may include multiple first data samples and a first sample label corresponding to each first data sample. The first sample label may indicate the fault type of the corresponding first data sample. The fault types indicated by the sample labels of the datasets described herein may include fault-free types and fault types with faults.

[0121] As an example, the first dataset can be obtained from vibration image data (including time-domain waveforms and spectrum diagrams) generated based on historical vibration data of wind turbine blades. Here, the processing procedure for obtaining vibration image data from historical vibration data is the same as the processing procedure for generating vibration image data described above, so it will not be repeated here.

[0122] It should be noted that the datasets mentioned in this article (such as the first dataset here and the second and third datasets described below) can include both faulty data samples and non-faulty data samples.

[0123] In step S1620, the second large model can be used to extract the fault features corresponding to each fault type based on the first dataset and generate initial prompt words.

[0124] In this step, the first dataset can be input into the second large model to extract fault features, generate the optimized part of the prompt words, and then splice or combine it with the preset fixed part to obtain the initial prompt words.

[0125] As an example, the core generation steps of the prompt word optimization part may include feature extraction and judgment criterion construction.

[0126] Here, feature extraction can refer to extracting the core time-domain and spectral features from the first dataset according to the fault type. The features must be quantifiable, identifiable, and unambiguous (e.g., avoid using "large amplitude" and instead use "amplitude suddenly increases to 20~30 m / s"). 2 The above). The judgment criteria construction can refer to classifying the extracted features according to "normal / abnormal + fault subtype," with each fault type corresponding to a unique combination of "time-domain features + spectral features." Below is an example of an optimized portion of the prompt words: text Judgment criteria correspond to fault types: 1. Normal (answer=0, fault_type="normal"): Time domain: The vibration of all three blades is within ±10 m / s of 0. 2 The range fluctuates, exhibiting regular and stable periodic vibrations as the impeller rotates, with a stable phase difference and no abnormal spikes or disturbances. Spectrum: Energy is concentrated at the impeller rotation frequency (1st harmonic) and 2nd to 3rd harmonics, with stable peak values ​​and no additional noise, broadband noise or discrete spikes. The spectral characteristics of each blade are consistent.

[0127] 2. Abnormal (answer=1): A condition is considered abnormal if it meets any of the following characteristics, and the corresponding fault type is matched: Blade cracks / damage (fault_type="blade cracks / damage") Time domain: Periodic impact spikes occur (the spike interval is synchronized with the impeller rotation frequency, approximately 3-6 seconds / time), with amplitude exceeding ±10m / s. 2 And it persists, accompanied by abrupt changes in vibrational phase; Spectrum: Sideband characteristics appear centered on the first harmonic (equally spaced frequency peaks on both sides, with the spacing consistent with the blade's passing frequency), or high-frequency modulation components are present (the spectrum is comb-shaped).

[0128] Quality imbalance (fault_type="quality imbalance") Time domain: The vibration amplitude of a single blade increases gradually (from ±10 m / s within 120 seconds). 2 Rise to ±25m / s 2 (Above), there was no obvious impact but it deviated from the normal fluctuation range and was contrary to the vibration trend of other blades; Spectrum: The amplitude of the impeller rotation frequency (1st harmonic) is abnormally high (more than 2 times the normal spectrum), the amplitude of the 2nd to 3rd harmonics drops sharply, and the energy distribution is unbalanced.

[0129] The above optimization section is just an example; other fault types can be supplemented according to this format.

[0130] In step S1620, the second large model can also be a multimodal large model with image and natural language understanding capabilities. The second large model can be the same as the first large model mentioned above, or it can be a different large model.

[0131] In step S1630, the first large model can be used to optimize the initial prompt words and determine the preset prompt words.

[0132] In this step, the effectiveness of the initial prompts can be verified using the first major model. Based on the verification results, it can be determined whether and how to optimize the prompts.

[0133] As an example, step S1630 may include: obtaining a second dataset; using the first large model and the second dataset, performing at least one round of prompt word optimization on the initial prompt words to determine the preset prompt words.

[0134] Here, the second dataset (also known as the test set) may include multiple second data samples and a second sample label corresponding to each second data sample, the second sample label indicating the fault type of the corresponding second data sample.

[0135] As an example, the first dataset mentioned above, the second dataset here, and the third dataset below (in the example where a third dataset exists) can be obtained by dividing the vibration image dataset generated from historical vibration data. Here, for the vibration image dataset generated from historical vibration data, all images can be named in a unified format (e.g., "fault type_acquisition time_time domain / spectral graph.jpg") to facilitate the accurate return of the "original image name" in the prompt.

[0136] For example, as shown in Table 1 below, the total set can be divided into three independent datasets according to the function and data volume of each dataset, and it is ensured that there is no sample overlap between the datasets.

[0137] Table 1

[0138] With the second dataset obtained, at least one round of prompt word optimization can be performed on the initial prompt words to determine the preset prompt words.

[0139] As an example, such as Figure 17 As shown, the above-mentioned prompt word optimization may include the following steps: In step S1710, the first large model can be used to perform fault diagnosis on the blade based on the prompt words to be optimized in this round and multiple second data samples to obtain test diagnosis results.

[0140] Specifically, in the initial prompt word optimization, the prompt words to be optimized in this round are the initial prompt words; in subsequent prompt word optimizations, the prompt words to be optimized in this round are the prompt words obtained from the previous round of optimization.

[0141] The prompts to be optimized in this round and multiple second data samples (e.g., all second data samples in the second dataset) can be input into the first large model. The first large model is then used for fault diagnosis to obtain test diagnosis results. Here, the test diagnosis results can, for example, have the same or similar structure and format as the fault diagnosis results described above.

[0142] In step S1720, a comparison result can be obtained by comparing the test diagnosis result with the second sample label.

[0143] In this step, the effectiveness of the current prompt word can be verified by comparing the test diagnosis results with the second sample label. Here, the comparison result characterizes the difference between the test diagnosis results and the second sample label.

[0144] As an example, the comparison results may include: precision metrics for each fault type; and / or, recall metrics for each fault type.

[0145] Here, the precision metric can be determined based on the number of samples correctly diagnosed as the fault type in multiple second data samples and the total number of samples diagnosed as the fault type; the recall metric can be determined based on the number of samples correctly diagnosed as the fault type in multiple second data samples and the total number of samples of the fault type, where the total number of samples of the fault type is determined, for example, by the second sample labels. For example, the precision and recall metrics can be represented as follows: Accuracy metric = Number of samples correctly diagnosed as this fault type / Total number of samples diagnosed as this fault type; Recall rate = Number of samples correctly diagnosed as this fault type / Total number of samples of this fault type.

[0146] Return to reference Figure 17 In step S1731, in response to the comparison result meeting the preset optimization conditions, a preset prompt word can be determined based on the prompt word to be optimized in this round; in step S1732, in response to the comparison result not meeting the preset optimization conditions, the prompt word to be optimized in this round can be optimized, and the optimized prompt word can be used as the prompt word to be optimized in the next round.

[0147] Here, the preset optimization conditions can be related to the aforementioned precision and / or recall metrics. For example, they can include: precision ≥ s1 and recall ≥ s2, where s1 and s2 represent preset thresholds that can be set and adjusted according to actual needs; for example, both s1 and s2 can be 80%. When the above conditions are met, the final suggestion words can be determined based on the suggestion words to be optimized in this round; when the above conditions are not met, the next round of optimization can begin.

[0148] Here, in step S1732, as Figure 18 As shown, the prompts to be optimized in this round can be optimized in the following ways: In step S1810, the target data samples that were missed or misjudged in the multiple second data samples in this round of fault diagnosis can be identified.

[0149] Here, by comparing the test diagnostic results and the second sample label, the target data samples that are currently missed or misjudged can be identified.

[0150] In step S1820, the first dataset can be updated using the target data sample and the corresponding second sample label to obtain the updated first dataset.

[0151] In this step, for example, the target data sample and the corresponding second sample label can be added to the current first dataset to obtain the updated first dataset.

[0152] In step S1830, the second large model can be used to optimize the current prompt words based on the updated first dataset to obtain the optimized prompt words for this round.

[0153] In one example, the updated first dataset and the current prompt words (e.g., the prompt words to be optimized in this round) can be input into the second large model to optimize the current prompt words using the second large model, thus obtaining the optimized prompt words for this round.

[0154] In another example, guidance information can be introduced when optimizing the current prompt.

[0155] For example, step S1830 above may include: generating guidance information based on the target data sample and the current prompt word; and using the second large model, optimizing the current prompt word based on the updated first dataset and the guidance information to obtain the optimized prompt word for this round.

[0156] Here, the guidance information can suggest the optimization direction for the second major model to optimize the current prompt word.

[0157] As an example, the step of generating guidance information based on the target data sample and the current prompt word may include: in response to the fact that the description of the fault type of the target data sample in the current prompt word has a preset defect, generating guidance information corresponding to the preset defect.

[0158] Here, the pre-defined defects may include at least one of the following: vague feature descriptions; overlapping diagnostic rules for different fault types; lack of diagnostic rules for fault edges; the guidance information may include at least one of the following: extracting quantitative features; supplementing differential features; adding diagnostic rules for fault edges.

[0159] Specifically, if misjudgment is caused by "vague feature description", the prompt words can be optimized by quantifying the features. For example, the vaguely described feature "abnormal peak value in high frequency band" can be changed to the quantified feature "amplitude above 200Hz exceeds 5 times the normal value".

[0160] If misjudgment is caused by "overlapping fault type features", the prompt words can be optimized by adding differentiated features. For example, in the fault type of structural loosening, "continuous occurrence of 4th harmonic and above" can be added to distinguish it from the fault type of loss of coordination.

[0161] If a misjudgment is caused by "missed edge samples", the prompt can be optimized by adding edge feature descriptions to the prompt. For example, for the fault type of mass imbalance, "amplitude rises to 20~25m / s within 120 seconds" can be added. 2 The critical situation of "".

[0162] Here, the optimized prompt words from this round will be used as input for the next round of prompt word optimization. The first large model and the second dataset will be used for testing again until the comparison results of the test diagnosis results meet the above-mentioned preset optimization conditions. Then, prompt words can be output or the process can enter the verification stage of the third dataset.

[0163] In addition, as an example, a test termination condition can be set, which may include the number of tests using the second dataset being greater than or equal to a preset number of tests to avoid over-optimization, such as the number of tests being ≥ m, where m can be a positive integer, such as 10.

[0164] Furthermore, as an example, in step S1731 above, when the preset optimization conditions are met, in one example, the prompt word to be optimized in this round can be directly determined as the final prompt word.

[0165] In another example, the first large model and the third dataset can be used to validate the prompt words to be optimized in this round, and the final prompt words can be determined based on the validation results.

[0166] Here, the third dataset (also known as the validation set) may include multiple third data samples and a third sample label corresponding to each third data sample. The third sample label indicates the fault type of the corresponding third data sample. The method for obtaining the third dataset has already been described above and will not be repeated here.

[0167] Specifically, the first large model and the prompt words after passing the test on the second dataset can be used to diagnose the faults in the third dataset, obtain the validation diagnosis results, and compare the validation diagnosis results with the third sample labels to obtain the comparison results for the validation diagnosis results (such as the aforementioned precision and / or recall metrics).

[0168] If the comparison results for the verification and diagnosis results meet the aforementioned preset optimization conditions, the current prompt word (e.g., the prompt word to be optimized in this round that has passed the test on the second dataset) can be determined as the preset prompt word, thereby terminating the optimization and outputting the final prompt word. If the comparison results for the verification and diagnosis results do not meet the preset optimization conditions, misjudged samples from the third dataset can be added to the first dataset, and the process of "optimizing the prompt word to be optimized in this round and using the optimized prompt word as the prompt word to be optimized in the next round" can be repeated. In this way, the process of "prompt word adjustment → test set verification → verification set verification" can be repeated until the prompt word meets the criteria, such as the comparison results for the verification and diagnosis results meeting the aforementioned preset optimization conditions.

[0169] Furthermore, as an example, an optimization termination condition can be set, which can be related to the number of optimization rounds to avoid infinite iteration. For example, the optimization termination condition could include: cumulative optimization count ≥ n, where n is a positive integer, such as 8.

[0170] Figure 19 A flowchart illustrating an application example of prompt word optimization according to an exemplary embodiment of the present disclosure is shown.

[0171] like Figure 19 As shown, in step S1910, the dataset can be divided to obtain the first dataset, the second dataset, and the third dataset mentioned above.

[0172] In step S1920, initial prompt words can be generated using the first dataset, which may include a fixed part and an optimized part.

[0173] In step S1930, the initial prompt words can be tested using the second dataset (or test set), and comparison results for the test diagnostic results, such as precision and / or recall metrics, can be calculated.

[0174] In step S1940, it can be determined whether the test indicators in the comparison results of the test diagnosis results meet the standards, such as whether the preset optimization conditions are met.

[0175] In response to the test metrics being met in step S1940, in step S1950, a final validation can be performed using a third dataset (or validation set), and comparison results for the validation diagnostic results, such as precision and / or recall metrics, can be calculated.

[0176] In step S1960, it can be determined whether the verification indicators in the comparison results of the verification diagnosis results meet the standards, such as whether the preset optimization conditions are met.

[0177] In response to the verification that the target has been met in step S1960, the final prompt word can be output in step S1970.

[0178] In response to the failure of the test index to meet the standard in step S1940 or the failure of the verification index to meet the standard in step S1960, in step S1980, samples that were missed or misjudged can be screened and added to the first dataset.

[0179] In step S1990, the current prompt word can be optimized, for example, the judgment criteria can be adjusted, and then return to step 1930 to enter the next round of testing and optimization.

[0180] Taking the prompt word example from the previous text as an example, the final output prompt word example after the above optimization mechanism is as follows: The `prompt_text` attribute in the text is a sequence diagram and spectrum of vibration waveforms from three blades of a wind turbine. It requires completion of "Normal / Abnormal Judgment," "Fault Type Classification," and "Confidence Level," and returns the results in a specified JSON format. No extra text, punctuation, code blocks, or formatting. Image names should directly use the original .jpg / .png endings. Analysis and descriptions should only extract the core judgment criteria and not expand upon them. Strictly adhere to this format: {"file":"Original Image Name","analysis":"Core Judgment Criteria (Time Domain + Spectral Characteristics)","answer":"0 / 1","fault_type":"Specific Fault Type / Normal","confidence":"Confidence Level"} Judgment criteria correspond to fault types: 1. Normal (answer=0, fault_type="normal"): Time domain: Vibration of all 3 blades is within ±10m / s of 0. 2The range fluctuates, exhibiting regular and stable periodic vibrations as the impeller rotates, with a stable phase difference and no abnormal spikes or disturbances; the spectrum: the energy is concentrated at the impeller rotation frequency (1st harmonic) and the 2nd to 3rd harmonics, with stable peak values ​​and no additional noise, broadband noise or discrete spikes, and the spectral characteristics of each blade are consistent.

[0181] 2. Abnormal (answer=1): A condition is considered abnormal if it meets any of the following characteristics, and the corresponding fault type is matched: Blade crack / damage (fault_type="blade crack / damage") time domain: periodic impact spikes occur (the spike interval is synchronized with the impeller frequency, about 3~6 seconds / time), amplitude exceeds ±10m / s 2 And it persists, accompanied by abrupt changes in vibration phase; spectrum: sideband characteristics appear centered on the first harmonic (equally spaced frequency peaks on both sides, with the interval consistent with the blade's passing frequency), or there are high-frequency modulation components (the spectrum is comb-shaped).

[0182] Mass imbalance (fault_type="mass imbalance") Time domain: The vibration amplitude of a single blade increases gradually (from ±10 m / s within 120 seconds). 2 Rise to ±25m / s 2 The above shows no obvious impact but deviates from the normal fluctuation range and is contrary to the vibration trend of other blades; Spectrum: The amplitude of the impeller rotation frequency (1st harmonic) is abnormally high (more than 2 times the normal spectrum), the amplitude of the 2nd to 3rd harmonics drops sharply, and the energy distribution is unbalanced.

[0183] Structural loosening (fault_type="structural loosening") Time domain: Disordered vibration coordination of multiple blades, irregular phase difference of the time domain waveforms of the three blades, no synchronous periodic characteristics, and chaotic cross-vibration curves; Spectrum: Abnormally high amplitude peaks appear in the high frequency band (more than 5 times the normal high frequency energy), or continuous occurrence of harmonics of the 4th harmonic and above (amplitude close to the 1st harmonic, dense peaks of multiple harmonics).

[0184] Sensor fault (fault_type="sensor fault") Time domain: The signal exhibits saturation truncation (amplitude remains constant at a certain threshold without fluctuation), disconnection blank segment (time domain curve is flat with no data, lasting for more than 5 seconds), or no data input; Spectrum: Irregular broadband noise appears (energy rises uniformly across the entire frequency band, with no obvious dominant frequency), or discrete clutter spikes (isolated high amplitude peaks at non-harmonic frequencies, exceeding the normal by more than 3 times).

[0185] Aerodynamic instability (fault type="aerodynamic instability (fault)" time domain: vibration amplitude suddenly increases to 20~30m / s 2The above shows that the amplitude fluctuates greatly, deviating from the periodic vibration characteristics, and the signal has irregular and chaotic jumps and discontinuous vibration curves; the spectrum shows that the energy in the low frequency band (0~5Hz) is significantly increased, accompanied by a shift in the main frequency (the frequency shifts with time by more than 0.5Hz), or there is no obvious peak in the wideband noise across the entire frequency band.

[0186] Time domain of failure of coordination (fault_type="blade failure"): Single blade vibration amplitude significantly exceeds the limit (flapping ≤ -40m / s) 2 1. Shaking vibration ≤ -30m / s 2 The spike pulses, spaced approximately 6 seconds apart and continuously occurring within a 120-second acquisition period, are completely out of sync with the vibrations of the other blades; Spectrum: The spectral characteristics of each blade differ significantly (amplitude difference exceeding 2 times at the first harmonic, and differences in the number / position of harmonic components), with the spectrum of a single blade being completely chaotic and irregular. Example explanation of prompt word parameters: ±10m / s 2 The threshold value is the 95% confidence interval based on statistics from more than 3,000 sets of normal blade vibration data.

[0187] Using the above methods, prompt words that conform to the operating mechanism of wind turbine generators can be constructed. Based on the core ideas of dataset partitioning, prompt word generation in fixed / optimized parts, and iterative optimization mechanism, a complete and practical prompt word generation and optimization system for blade large model fault diagnosis algorithm can be constructed. Thus, through iterative optimization, the prompt words can be accurately adapted to the fault diagnosis task, improving the classification accuracy and recall rate.

[0188] The above approach, centered on "data-driven" and "iterative optimization," employs a hierarchical dataset, modular generation of prompt words, and multiple rounds of verification iterations to ultimately obtain high-quality prompt words adapted to fault diagnosis of wind turbine blade vibration data. Furthermore, the approach, with its core process of "test set verification → supplementing misjudged samples → adjusting prompt words → final verification on the validation set," ensures the accuracy and generalization of the prompt words.

[0189] Furthermore, in embodiments of this disclosure, the blade fault diagnosis method may further include: acquiring multiple fault diagnosis results output by a large model; aggregating the trends of the multiple fault diagnosis results to obtain an aggregation result; and determining the confidence level of the fault diagnosis result based on the aggregation result.

[0190] Specifically, for example, a time-sliding window mechanism can be used to confirm and aggregate trends of multiple fault diagnosis results output by the first large model, thereby avoiding false alarms caused by instantaneous data anomalies or fluctuations in a single inference by the large model. Here, the window length of the sliding window can be, for example, 30 minutes, and each time two days of historical calculation results from the large model are retrieved for comprehensive analysis.

[0191] The trends of multiple fault diagnosis results can be aggregated based on preset aggregation rules. For example, multiple fault diagnosis results returned by the first large model can be aggregated and analyzed within a sliding window. A valid abnormal event is considered to have occurred only when the number of times an anomaly (e.g., "answer=1") is met within the window exceeds a preset proportion (e.g., 50%) and the main fault type (fault_type) is consistent.

[0192] When the aggregated results indicate that an abnormal event has occurred, the confidence levels can be aggregated. For example, the average confidence level of the same fault type within the window can be calculated as the overall confidence level of this abnormal event.

[0193] Furthermore, in the embodiments of this disclosure, the blade fault diagnosis method may further include: in response to a fault diagnosis result indicating the presence of a target fault type, performing fault diagnosis on the blade based on vibration data using a preset mechanism model to obtain a verification diagnosis result; in response to a verification diagnosis result indicating the presence of a target fault type, outputting a fault diagnosis result; and in response to a verification diagnosis result indicating the absence of a target fault type, updating the preset prompt words based on vibration data and the verification diagnosis result.

[0194] Here, the mechanism model can be a model that uses the operating principle of wind turbine generators to diagnose the target fault type. For example, the mechanism model can be a statistical diagnostic model based on time domain characteristics and / or frequency domain characteristics.

[0195] As an example, the mechanistic model can be viewed as a smaller model, which can be set up separately for a single fault type. The mechanistic model can be used to verify the output of the first large model, or in other words, to perform a secondary verification of the large model's diagnostic results. By verifying the mechanistic model, the false alarm rate can be reduced, especially for misjudging serious faults such as "blade cracks / damage," "loss of coordination," and "aerodynamic instability."

[0196] As an example, for the six fault types output by the large model, a dedicated lightweight mechanistic model can be constructed for each. The input of each mechanistic model is the vibration data of the blade; the output of each mechanistic model is the verification and diagnosis result, which can be, for example, a conclusion of "diagnosis consistent" or "diagnosis corrected". Here, examples of the mechanistic models corresponding to the six fault types are shown in Table 2 below.

[0197] Table 2

[0198]

[0199] In response to a fault indicating the presence of a target fault type in the verification diagnosis result, such as a "diagnosis consistent" result, the large model diagnosis can be deemed valid, and the fault diagnosis result can be output. For example, the abnormal event can be pushed to the hierarchical alarm process described below.

[0200] In response to a fault where the review diagnosis result indicates that the target fault type does not exist, such as a "diagnosis correction" result, the fault type and the analysis of the large model can be corrected based on the mechanism model (e.g., "blade crack / damage" can be corrected to "sensor fault"). This revised example (including the original vibration data, the large model result, and the correction result) is then added to the first dataset to update the prompt words.

[0201] As an example, in response to the number of amendment examples added to the first dataset reaching a preset number (e.g., 20), the prompt word optimization mechanism can be automatically triggered (e.g., as mentioned above). Figures 16 to 19 (The process described).

[0202] Furthermore, in embodiments of this disclosure, the blade fault diagnosis method may further include: executing a graded alarm strategy based on the fault type (fault_type) and average confidence level of the aggregated abnormal events.

[0203] Here, a tiered alarm strategy can include multiple triggering conditions and multiple execution actions corresponding to each triggering condition. Different triggering conditions can correspond to different alarm levels. Alarm levels can include multiple levels.

[0204] As an example, the alarm level can include a Level 1 alarm, at which observation suggestions can be provided. Specifically, the triggering conditions include a first triggering condition and a first action corresponding to the Level 1 alarm. The first triggering condition may include, for example, a relatively slow-developing fault such as a fault type of "sensor failure" or "mass imbalance," with an average confidence level within a first range (e.g., 60%~80%). The first action may include, for example, marking the fault in yellow on the monitoring system interface and generating an observation log. Additionally, notifications can be pushed to the operations and maintenance personnel's application (APP).

[0205] As an example, alarm levels can include level 2 alarms, at which preventative maintenance can be provided. Specifically, triggering conditions include a second triggering condition and a second action corresponding to the level 2 alarm. The second triggering condition could include, for example, faults with a risk of development such as "blade crack / damage" or "structural loosening," and an average confidence level in the second interval (e.g., above 80%). The second action could include, for example, the system automatically generating a preventative maintenance work order, with the alarm information accompanied by the "core judgment criteria" generated by the first major model (e.g., the "analysis" field in the output of the first major model), clearly guiding the key points of operation and maintenance.

[0206] As an example, alarm levels can include three levels, at which emergency actions can be provided. Specifically, triggering conditions include a third triggering condition and a third action corresponding to a level three alarm. The third triggering condition could include, for example, a serious fault type such as "aerodynamic instability (flutter)" or "loss of coordination" that could trigger a chain reaction or safety accident, which would be triggered immediately regardless of confidence level. The third action could include, for example, triggering the highest-level audible and visual alarm and sending an emergency SMS / phone call to the responsible person. The work order is automatically escalated to an emergency work order, and emergency operations such as remote shutdown or power-limited operation are recommended.

[0207] Furthermore, in embodiments of this disclosure, the blade fault diagnosis method may further include: recording abnormal event data identified as abnormal events into a first dataset.

[0208] Specifically, the processing of all alarm events and the final confirmed data type and label can be recorded in the first dataset to achieve feedback loop and optimize prompt words.

[0209] As an example, the above-mentioned anomaly event data may include false alarm analysis data and / or false negative analysis data.

[0210] Specifically, false alarm analysis data can be obtained by analyzing cases (false alarms) where the system alarms but are confirmed to be fault-free after manual inspection. It indicates whether the business mechanism prompts used by the first model at that time were inappropriate or whether the vibration feature image was generated incorrectly.

[0211] Missed detection analysis data can be obtained by retrospectively analyzing cases where no alarms were triggered but subsequent failures occurred (missed detections). This data indicates whether the failure was due to the judgment rules in the alert words (e.g., the threshold "±10m / s"). 2 Inaccurate or missing features led to underreporting.

[0212] Furthermore, similar to the prompt word optimization mechanism mentioned above, the prompt word optimization mechanism can be automatically triggered when the number of abnormal event data added to the first dataset reaches a preset number (e.g., 20). Figures 16 to 19 (The process described).

[0213] According to the blade fault diagnosis method of the embodiments of this disclosure, a three-stage collaborative architecture can be implemented. In the first stage, the edge-side small model can complete the operating condition screening (e.g., removing shutdown and / or yaw data) and the "spectral ski slope" algorithm quality filtering, outputting high signal-to-noise ratio vibration data. In the second stage, the multimodal large model combines time-frequency images and business mechanism prompts to output structured diagnostic results through cross-modal reasoning. In the third stage, the dedicated mechanism small model reviews the results of the large model, corrects misjudgments, and supplements the case set to optimize prompts. The three stages here have clear division of labor, achieving both cost reduction and improved diagnostic accuracy.

[0214] Specifically, this blade fault diagnosis method achieves a synergistic breakthrough in diagnostic performance. Specifically, a lightweight, conventional front-end model (small model) efficiently performs data condition filtering (e.g., filtering invalid states like shutdown / yaw using SCADA data) and deep quality filtering (e.g., filtering out signal jumps and abnormal noise), ensuring high-quality input features from the source. The back-end, combined with a carefully designed business mechanism prompt word project, structurally encodes the physical and business knowledge of blade faults (e.g., fault type, key frequency band, feature thresholds) into text descriptions, guiding the large model after multimodal generation (time-frequency diagram) to perform accurate cross-modal analysis and reasoning. This collaborative framework effectively integrates the small model's ability to process industrial data mechanistically with the large model's powerful pattern diagnosis and complex logical reasoning capabilities, thereby achieving a synergistic improvement in early fault detection rate and overall diagnostic accuracy (reducing false positives).

[0215] Based on practical application verification, compared with traditional vibration signal diagnostic models, this method improves the recall rate of early cracks by more than 30%; compared with directly applying general multimodal large models, its accuracy is improved by more than 20%.

[0216] Furthermore, this blade fault diagnosis method also achieves high efficiency in computing resources and feasible deployment, forming an efficient resource collaboration paradigm where "small models process massive amounts of raw data, while large models focus on high-quality feature inference." This design significantly reduces the need for raw, low-quality data processing for computationally intensive large AI models, concentrating the computational load of large models on the core high-value diagnostic inference stage. The overall system's computing resource consumption is reduced by approximately 30%, making it more suitable for resource-constrained environments such as industrial edge computing units or field servers, and solving the engineering challenge of real-time, large-scale deployment of large models in cost-sensitive scenarios.

[0217] Furthermore, this blade fault diagnosis method enhances the interpretability and engineering applicability of the diagnostic process. Specifically, by introducing structured business mechanism hints, the diagnostic decision-making process of the AI ​​large model becomes transparent and traceable. The output diagnostic results are accompanied by domain-knowledge-based logical analysis (such as the "analysis" field mentioned above), significantly improving the engineering interpretability of the diagnostic conclusions and the trustworthiness of maintenance personnel, thus facilitating subsequent maintenance decisions.

[0218] Furthermore, this blade fault diagnosis method also enables the system to achieve self-adaptation and continuous evolution. Specifically, the method incorporates a closed-loop optimization mechanism based on operational feedback. By analyzing false alarms and missed alarms, the judgment rules and feature thresholds of business prompt words can be iteratively optimized, and the optimized prompt words and newly added fault case data are used to continuously enhance the system. This allows the system to gradually adapt to new or complex faults, forming a continuous evolutionary closed loop of "data-diagnosis-feedback-optimization," improving the robustness and accuracy of long-term applications.

[0219] Figure 20 An example framework for implementing this blade fault diagnosis method is shown. Figure 20 As shown, the data input module 2001 can input data from SCADA data sources and blade vibration data sources. The operating condition screening module 2002 can screen the vibration data based on operating conditions, and the data filtering module 2003 can filter the data. Here, data filtering may include, for example, data filtering of conventional models and depth data filtering of the aforementioned spectrum skid slope algorithm. The vibration image data generation module 2004 can generate vibration image data based on the filtered vibration data. The large model diagnosis and small model verification module 2005 can diagnose faults using a large model (the aforementioned first large model) and can verify the diagnosis results of the large model using a small model (the aforementioned mechanism model). The result output and application module 2006 can output and apply the verified results. The result display module 2007 can display the final fault diagnosis results, such as displaying alarms on a screen. Furthermore, the modules 2001 to 2006 can be implemented on a central processing unit (CPU), and the modules 2001 to 2006 can be implemented on a graphics processing unit (GPU).

[0220] As an example, the modules listed in 2001 to 2007 above can be distributed across the edge computing units of wind turbine generators, site servers, or cloud platforms, and can be configured according to specific resource availability.

[0221] The blade fault diagnosis method according to the exemplary embodiments of this disclosure can effectively integrate the data cleaning and mechanistic feature extraction capabilities of conventional models at the front end with the powerful pattern recognition and multimodal reasoning capabilities of large models at the back end, thereby synergistically improving the accuracy, recall, and early warning capabilities of blade fault identification, while optimizing computational resource consumption.

[0222] Specifically, this method utilizes lightweight conventional models to quickly and effectively filter and correct quality issues in the original vibration data (such as signal jumps, abnormal noise, and sensor failures), providing high-quality input for the upper-level model.

[0223] Furthermore, in this method, the processed high-quality vibration signal can be intelligently transformed into an image representation (e.g., time-spectrum graph) containing rich fault features, and through carefully designed cue word engineering, the operational mechanism knowledge of blade faults (e.g., fault type, key frequency band, feature value threshold) can be encoded into a text description and fused with the large model.

[0224] Furthermore, this method can also construct an efficient and collaborative diagnostic framework for large and small models, enabling the large model to accurately understand and utilize vibration data image features and business prompts, thereby significantly improving the overall recognition rate and recall rate of various faults such as blade cracks, imbalance, icing, and loose bolts.

[0225] By dividing tasks as described above, the requirements for raw data processing and real-time inference of large AI models are reduced. While ensuring diagnostic performance, the overall system's computing resource consumption is effectively controlled, making it more suitable for deployment in industrial settings.

[0226] Furthermore, according to the exemplary embodiment of the blade fault diagnosis method disclosed herein, a conventional model is responsible for front-end data governance (e.g., operating condition screening, quality filtering), while an AI big model is responsible for back-end accurate reasoning. The two are seamlessly connected through a combination of high-quality data and structured prompts, which solves the problems of ambiguous division of labor and insufficient collaboration in related technologies. The responsibilities of data quality governance and advanced intelligent diagnosis are clearly defined, and the optimal balance between performance and efficiency is achieved.

[0227] According to the exemplary embodiments of the present disclosure, the business mechanism prompt word engineering in the blade fault diagnosis method can structure and encode professional knowledge such as fault judgment rules and frequency band thresholds, enabling the AI ​​large model to directly understand the business mechanism and breaking through the bottleneck of general large models lacking professional knowledge. Inputting the professional knowledge and rules of fault diagnosis into the large model in the form of structured natural language as a guided and interpretable context is the core technology of this solution for improving the diagnostic accuracy in specialized fields.

[0228] The blade fault diagnosis method according to the exemplary embodiments of this disclosure can also realize resource-optimized collaborative diagnosis. The design of the whole scheme fully considers the industrial implementation cost, reduces the overall computational complexity through division of labor, and achieves a balance between high performance and feasibility.

[0229] A second aspect of the embodiments of this disclosure provides a blade fault diagnosis device based on a large model, the generation device including an acquisition unit, a generation unit, and a diagnosis unit.

[0230] The acquisition unit is configured to acquire vibration data of the blades of a wind turbine generator.

[0231] The generation unit is configured to generate vibration image data based on vibration data, wherein the vibration image data includes a time-domain plot and / or a spectrogram.

[0232] The diagnostic unit is configured to use the first large model to perform fault diagnosis on the blade based on preset prompts and vibration image data, and obtain fault diagnosis results. The preset prompts indicate the diagnostic rules for fault diagnosis performed by the first large model.

[0233] As an example, the generation unit is configured to: perform data filtering on the vibration data to obtain filtered vibration data; generate vibration image data based on the filtered vibration data, wherein the vibration data includes multiple time-series vibration data, and the data filtering includes: converting each time-series vibration data into a vibration spectrum; and filtering the multiple time-series vibration data according to the changing trend of the amplitude of each vibration spectrum to obtain filtered time-series vibration data.

[0234] As an example, the generation unit is configured to: perform envelope fitting on the amplitude of each vibration spectrum to obtain the amplitude boundary; for each vibration spectrum, determine the data change trend of the amplitude boundary of the vibration spectrum within each moving window by moving the frequency window multiple times on the vibration spectrum; determine whether the data quality of the time series vibration data is abnormal based on the consistency of the data change trend within the multiple moving windows; and obtain filtered time series vibration data by removing data with abnormal data quality from multiple time series vibration data.

[0235] As an example, the generation unit is configured to: in response to the fact that the data change trend of the target window in the window that has been moved multiple times is all downward and the effective value of the data in the preset frequency range is greater than the preset value, determine that the data quality of the time series vibration data is abnormal, wherein the target window is a window with an upward or downward data change trend.

[0236] As an example, the blade fault diagnosis device also includes a verification unit, which is configured to: in response to a fault diagnosis result indicating the presence of a target fault type, perform fault diagnosis on the blade based on vibration data using a preset mechanism model to obtain a verification diagnosis result, wherein the mechanism model is a model for diagnosing the target fault type using the operating principle of a wind turbine generator; in response to a verification diagnosis result indicating the presence of a target fault type, output a fault diagnosis result; and in response to a verification diagnosis result indicating the absence of a target fault type, update preset prompt words based on vibration data and the verification diagnosis result.

[0237] As an example, the blade fault diagnosis device also includes a prompt word determination unit, which is configured to determine preset prompt words by: acquiring a first dataset, wherein the first dataset includes multiple first data samples and a first sample label corresponding to each first data sample, the first sample label indicating the fault type of the corresponding first data sample; using a second large model, based on the first dataset, extracting the fault features corresponding to each fault type and generating initial prompt words; and using the first large model, optimizing the initial prompt words to determine the preset prompt words.

[0238] As an example, the prompt word determination unit is configured to: acquire a second dataset, wherein the second dataset includes multiple second data samples and a second sample label corresponding to each second data sample, the second sample label indicating the fault type of the corresponding second data sample; perform at least one round of prompt word optimization on the initial prompt words using a first large model and the second dataset to determine preset prompt words, wherein the prompt word optimization includes: using the first large model, performing fault diagnosis on the blade based on the prompt words to be optimized in this round and multiple second data samples to obtain test diagnosis results; obtaining comparison results by comparing the test diagnosis results with the second sample labels; determining preset prompt words based on the prompt words to be optimized in this round in response to the comparison results meeting preset optimization conditions; and optimizing the prompt words to be optimized in this round in response to the comparison results not meeting the preset optimization conditions, and using the optimized prompt words as the prompt words to be optimized in the next round.

[0239] As an example, the prompt word determination unit is configured to optimize the prompt words to be optimized in this round by: identifying the target data samples that were missed or misjudged in multiple second data samples in this round of fault diagnosis; updating the first dataset using the target data samples and the corresponding second sample labels to obtain the updated first dataset; and using the second large model, optimizing the current prompt words based on the updated first dataset to obtain the optimized prompt words for this round.

[0240] As an example, the prompt word determination unit is configured to: generate guidance information based on the target data sample and the current prompt word; and optimize the current prompt word using the second large model based on the updated first dataset and the guidance information to obtain the optimized prompt word for this round, wherein the guidance information indicates the optimization direction of the second large model for optimizing the current prompt word.

[0241] As an example, the prompt word determination unit is configured to: in response to a preset defect in the description of the fault type of the target data sample in the current prompt word, generate guidance information corresponding to the preset defect, wherein the preset defect includes at least one of the following: ambiguous feature description; overlapping diagnostic rules for different fault types; lack of diagnostic rules for fault edges, and the guidance information includes at least one of the following: extracting quantitative features; supplementing differential features; adding diagnostic rules for fault edges.

[0242] As an example, the comparison results include: a precision metric for each fault type; and / or a recall metric for each fault type, wherein the precision metric is determined based on the number of samples correctly diagnosed as the fault type in a plurality of second data samples and the total number of samples diagnosed as the fault type; and the recall metric is determined based on the number of samples correctly diagnosed as the fault type in a plurality of second data samples and the total number of samples of the fault type.

[0243] Regarding the apparatus in the above embodiments, the specific manner in which each unit performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0244] A third aspect of the embodiments of this disclosure provides an electronic device, the electronic device including: a processor; and a memory for storing processor-executable instructions, wherein the processor-executable instructions, when executed by the processor, cause the processor to perform a blade fault diagnosis method based on a large model according to embodiments of this disclosure.

[0245] As an example, an electronic device does not necessarily have to be a single device; it can also be a collection of any means or circuits capable of executing the aforementioned instructions (or instruction sets) individually or in combination. An electronic device can also be part of an integrated control system or system manager, or can be configured to interface with a server, either locally or remotely (e.g., via wireless transmission).

[0246] In electronic devices, a processor may include, for example, a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example and not limitation, a processor may also include analog processors, digital processors, microprocessors, multi-core processors, processor arrays, network processors, etc.

[0247] The processor can execute instructions or code stored in memory, which can also store data. Instructions and data can also be sent and received over a network via a network interface device, which can employ any known transport protocol.

[0248] Memory can be integrated with the processor; for example, RAM or flash memory can be housed within an integrated circuit microprocessor. Alternatively, memory can comprise a separate device, such as an external disk drive, storage array, or other storage device that can be used by any database system. Memory and processor can be operatively coupled, or can communicate with each other, for example, via I / O ports, network connections, etc., enabling the processor to read files stored in the memory.

[0249] In addition, electronic devices may include video displays (such as liquid crystal displays) and user interaction interfaces (such as keyboards, mice, touch input devices, etc.). All components of the electronic device can be interconnected via buses and / or networks.

[0250] A fourth aspect of the embodiments of this disclosure provides a wind turbine generator set, the wind turbine generator set including electronic equipment as described in the embodiments of this disclosure, or the wind turbine generator set being communicatively connected to electronic equipment as described in the embodiments of this disclosure.

[0251] A fifth aspect of the embodiments of this disclosure provides a computer-readable storage medium that, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform a large-model-based blade fault diagnosis method according to embodiments of this disclosure.

[0252] Specifically, the blade fault diagnosis method based on a large model according to embodiments of this disclosure can be programmed into a computer program and stored on a computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by at least one processor, the at least one processor is caused to perform the blade fault diagnosis method based on a large model according to exemplary embodiments of this disclosure. Examples of computer-readable storage media include: read-only memory (ROM), random access programmable read-only memory (PROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), dynamic random access memory (DRAM), static random access memory (SRAM), flash memory, non-volatile memory, CD-ROM, CD-R, CD+R, CD-RW, CD+RW, DVD-ROM, DVD-R, DVD+R, DVD-RW, DVD+RW, DVD-RAM, BD-ROM, BD-R, BD-RLTH, B D-RE, Blu-ray or optical disc storage, hard disk drive (HDD), solid-state drive (SSD), card storage (such as multimedia cards, secure digital (SD) cards, or ultra-fast digital (XD) cards), magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, and any other device configured to store computer programs and any associated data, data files, and data structures in a non-transitory manner and to provide the computer programs and any associated data, data files, and data structures to a processor or computer so that the processor or computer can execute the computer programs. In one example, the computer programs and any associated data, data files, and data structures are distributed across a networked computer system, such that the computer programs and any associated data, data files, and data structures are stored, accessed, and executed in a distributed manner through one or more processors or computers.

[0253] A sixth aspect of the embodiments of this disclosure provides a computer program product including computer-executable instructions that, when executed by at least one processor, implement the large-model-based blade fault diagnosis method according to the embodiments of this disclosure.

[0254] According to the embodiments of this disclosure, a large-model-based blade fault diagnosis scheme constructs a full-process composite collaborative architecture from front-end processing, intermediate diagnosis, to back-end verification. In the first stage (small model pre-filtering), a lightweight conventional model deployed at the edge can efficiently perform condition screening (based on SCADA data) and preliminary quality filtering (filtering out noise, jumps, and other interference signals) on the raw vibration data. Then, the "spectral ski slope" algorithm is applied for deep filtering to obtain high-quality data, which is transformed into a spectrum image (such as a time-frequency graph) rich in fault features, providing a foundation for subsequent diagnosis. In the second stage (large model collaborative diagnosis), a multimodal AI large model that supports text and image question answering can be used, and business mechanism prompts are integrated as guidance. The large model receives the spectrum image and prompts output from the previous stage, performs cross-modal analysis and intelligent reasoning, and provides a preliminary diagnostic conclusion including fault type, severity, and confidence level. In the third stage (small model verification), the preliminary conclusions output by the large model can be verified using another (or more) lightweight, high-precision small models to correct any illusionary errors that may exist in the large model, thus forming the final diagnostic results and ensuring high reliability.

[0255] Furthermore, according to the large-model-based blade fault diagnosis scheme of the embodiments of this disclosure, a business mechanism prompt word engineering method and optimization mechanism for wind turbine blade fault diagnosis are provided, proposing a method to structure and encode domain knowledge into understandable and optimizable prompt words. This method first converts fault mechanisms (e.g., impact characteristics caused by cracks), judgment rules (energy thresholds for specific frequency bands), and engineering experience (spectral changes under different wind speeds) into instruction templates that can be parsed by the large model. By constructing three datasets—a case set, a test set, and a validation set—a dynamic update mechanism based on diagnostic performance indicators (e.g., precision and recall) is established: when the system's precision for a specific fault is lower than a set threshold, typical false alarm samples from the case set are automatically called to strengthen the discriminative feature descriptions in the prompt words; when the recall rate is not met, insufficiently covered fault features are extracted from the fault knowledge base and added to the prompt word condition judgment logic. In addition, real-time integration of operation and maintenance feedback corrects the algorithm results, forming a "warning-analysis-optimization" closed loop, enabling the system to maintain high robustness while possessing the ability for continuous adaptive evolution.

[0256] Furthermore, the blade fault diagnosis scheme based on a large model according to embodiments of this disclosure also provides a resource-optimized collaborative diagnosis and hierarchical decision-making mechanism. Based on a three-stage collaborative architecture of "small model pre-filtering - large model collaborative diagnosis - small model verification," the system achieves an optimal balance between computational efficiency and diagnostic accuracy. A small pre-filtering model deployed at the edge is responsible for early data quality governance, concentrating the computationally intensive large model inference load in the high-quality feature interpretation stage, reducing overall system computational consumption by approximately 30%. Simultaneously, it combines a time-sliding window to aggregate and analyze multiple diagnostic results, and the front-end "small model verification" mechanism further filters out potential over-inference biases in the large model. Finally, based on diagnostic confidence and fault severity, a three-level hierarchical alarm (e.g., observation / recommendation / emergency) is implemented, ensuring an early crack recall rate improvement of over 30% while increasing accuracy by over 20%, truly resolving the core contradiction between high-precision diagnosis and limited computing resources in industrial scenarios.

[0257] The description in this disclosure has been presented for illustrative purposes and is not intended to be exhaustive or limited to this disclosure. Many modifications, variations, and alternative embodiments will be apparent to those skilled in the art from the teachings presented in the foregoing description and the associated drawings.

[0258] Furthermore, it should be noted that although several examples of each step have been described above with reference to specific accompanying drawings, it should be understood that the embodiments of this disclosure are not limited to the combinations given in the examples. Steps appearing in different drawings can be combined, and the execution order of each step can be changed, without exhaustive examples here. Unless otherwise specifically stated, the order of steps in the method according to this disclosure is intended to be illustrative only, and the steps in the method according to this disclosure are not limited to the order specifically described above, but can be changed according to actual circumstances. In addition, at least one step in the method according to this disclosure can be adjusted, combined, or deleted according to actual needs.

[0259] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

[0260] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

[0261] The specific embodiments of this disclosure have been described in detail above. Although some embodiments have been shown and described, those skilled in the art should understand that modifications and variations can be made to these embodiments without departing from the principles and spirit of this disclosure, which are defined by the claims and their equivalents. Such modifications and variations should also be within the protection scope of the claims of this disclosure.

Claims

1. A blade fault diagnosis method based on a large model, characterized in that, The blade fault diagnosis method includes: To acquire vibration data of wind turbine blades; Based on the vibration data, vibration image data is generated, wherein the vibration image data includes a time-domain plot and / or a spectrum plot; Using the first large model, based on preset prompts and the vibration image data, fault diagnosis is performed on the blade to obtain fault diagnosis results. The preset prompts indicate the diagnostic rules for fault diagnosis performed by the first large model.

2. The blade fault diagnosis method according to claim 1, characterized in that, The generation of vibration image data based on the vibration data includes: The vibration data is filtered to obtain filtered vibration data; Based on the filtered vibration data, the vibration image data is generated. The vibration data includes multiple time-series vibration data, and the data filtering includes: Convert each time-series vibration data into a vibration spectrum; Based on the variation trend of the amplitude of each vibration spectrum, the multiple time-series vibration data are filtered to obtain filtered time-series vibration data.

3. The blade fault diagnosis method according to claim 2, characterized in that, The step of filtering the multiple time-series vibration data based on the amplitude variation trend of each vibration spectrum to obtain filtered time-series vibration data includes: Envelope fitting is performed on the amplitude of each vibration spectrum to obtain the amplitude boundary; For each vibration spectrum, the amplitude boundary of the vibration spectrum is determined by moving the frequency window multiple times within each moving window to observe the data change trend. Based on the consistency of the data change trends within the multiple moving windows, determine whether the data quality of the time-series vibration data is abnormal; The filtered time-series vibration data is obtained by removing data with abnormal quality from the multiple time-series vibration data.

4. The blade fault diagnosis method according to claim 3, characterized in that, The step of determining whether the data quality of the time-series vibration data is abnormal based on the consistency of the data change trends within the multiple moving windows includes: If the data change trend of the target window in the multiple moving windows is all downward and the effective value of the data within the preset frequency range is greater than the preset value, it is determined that the data quality of the time-series vibration data is abnormal, wherein the target window is a window with an upward or downward data change trend.

5. The blade fault diagnosis method according to claim 1, characterized in that, The blade fault diagnosis method also includes: In response to the fault diagnosis result indicating the presence of a target fault type, the blade is diagnosed based on the vibration data using a preset mechanism model to obtain a verification diagnosis result. The mechanism model is a model that uses the operating principle of a wind turbine generator to diagnose the target fault type. In response to the verification diagnosis result indicating the existence of a fault of the target fault type, the fault diagnosis result is output; In response to the verification diagnosis result indicating that there is no fault of the target fault type, the preset prompt word is updated based on the vibration data and the verification diagnosis result.

6. The blade fault diagnosis method according to claim 1, characterized in that, The preset prompt words are determined in the following way: Obtain a first dataset, wherein the first dataset includes multiple first data samples and a first sample label corresponding to each first data sample, and the first sample label indicates the fault type of the corresponding first data sample; Using the second model, based on the first dataset, extract the fault features corresponding to each fault type and generate initial prompt words; Using the first large model, the initial prompt words are optimized to determine the preset prompt words.

7. The blade fault diagnosis method according to claim 6, characterized in that, The step of optimizing the initial prompt words using the first large model to determine the preset prompt words includes: Obtain a second dataset, wherein the second dataset includes multiple second data samples and a second sample label corresponding to each second data sample, the second sample label indicating the fault type of the corresponding second data sample; Using the first large model and the second dataset, perform at least one round of prompt word optimization on the initial prompt words to determine the preset prompt words. The optimization of prompt words includes: Using the first large model, based on the prompt words to be optimized in this round and the multiple second data samples, the blade is diagnosed to obtain the test diagnosis results; The comparison result is obtained by comparing the test diagnosis result with the second sample label; In response to the comparison result satisfying the preset optimization conditions, the preset prompt words are determined based on the prompt words to be optimized in this round; In response to the comparison result not meeting the preset optimization conditions, the prompt words to be optimized in this round are optimized, and the optimized prompt words are used as the prompt words to be optimized in the next round.

8. The blade fault diagnosis method according to claim 7, characterized in that, The prompts to be optimized in this round will be optimized using the following methods: Identify the target data samples that were missed or misjudged in the plurality of second data samples during this round of fault diagnosis; The first dataset is updated using the target data sample and the corresponding second sample label to obtain the updated first dataset. Using the second large model, based on the updated first dataset, the current prompt words are optimized to obtain the optimized prompt words for this round.

9. The blade fault diagnosis method according to claim 8, characterized in that, The second large model is used to optimize the current prompt words based on the updated first dataset, resulting in optimized prompt words for this round, including: Based on the target data sample and the current prompt words, generate guidance information; Using the second major model, based on the updated first dataset and the guidance information, the current prompt words are optimized to obtain the optimized prompt words for this round. The guidance information indicates the optimization direction for the second major model to optimize the current prompt word.

10. The blade fault diagnosis method according to claim 9, characterized in that, The step of generating guidance information based on the target data sample and the current prompt words includes: In response to a preset defect in the description of the fault type of the target data sample in the current prompt word, guidance information corresponding to the preset defect is generated. The preset defects include at least one of the following: ambiguous feature descriptions; overlapping diagnostic rules for different fault types; lack of diagnostic rules for fault edges. The guidance information includes at least one of the following: extracting quantitative features; supplementing differential features; and adding diagnostic rules for fault edges.

11. The blade fault diagnosis method according to claim 7, characterized in that, The comparison results include: a precision metric for each fault type; and / or a recall metric for each fault type, wherein the precision metric is determined based on the number of samples correctly diagnosed as the fault type and the total number of samples diagnosed as the fault type in the plurality of second data samples; and the recall metric is determined based on the number of samples correctly diagnosed as the fault type and the total number of samples of the fault type in the plurality of second data samples.

12. A blade fault diagnosis device based on a large model, characterized in that, The blade fault diagnosis device includes: The acquisition unit is configured to acquire vibration data of the blades of a wind turbine generator set; The generation unit is configured to generate vibration image data based on the vibration data, wherein the vibration image data includes a time-domain plot and / or a spectrum plot; The diagnostic unit is configured to use a first large model to perform fault diagnosis on the blade based on preset prompts and the vibration image data, and obtain fault diagnosis results, wherein the preset prompts indicate the diagnostic rules for the first large model to perform fault diagnosis.

13. An electronic device, characterized in that, The electronic device includes: processor; Memory used to store processor-executable instructions. Wherein, when the processor-executable instructions are executed by the processor, they cause the processor to execute the blade fault diagnosis method based on a large model according to any one of claims 1 to 11.

14. A wind turbine generator set, characterized in that, The wind turbine generator set includes the electronic equipment according to claim 13, or the wind turbine generator set is communicatively connected to the electronic equipment according to claim 13.

15. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are executed by the processor of the electronic device, the electronic device is able to perform the blade fault diagnosis method based on a large model according to any one of claims 1 to 11.

16. A computer program product comprising computer-executable instructions, characterized in that, When the computer-executable instructions are executed by at least one processor, they implement the blade fault diagnosis method based on a large model according to any one of claims 1 to 11.

Images