A fan blade fault diagnosis method based on voiceprint recognition

Abstract

The application provides a fan blade fault diagnosis method based on voiceprint recognition, comprising: deploying sound collection equipment at multiple key positions of a wind turbine generator, synchronously collecting a blade rotation azimuth angle signal, and synchronously cutting and framing the collected original sound signal in units of blade rotation periods to obtain a framed sound signal; performing adaptive noise suppression and rotation synchronization enhancement processing on the framed sound signal, extracting multi-level voiceprint features, and encoding the voiceprint features into a unified multi-modal feature vector; inputting the multi-modal feature vector into a large language model that is fine-tuned in the field, using the audio understanding and semantic reasoning capabilities of the large language model to identify the type and severity of blade damage; outputting a structured diagnosis result containing the damage type, confidence, severity classification, explainability explanation and operation and maintenance suggestion, and storing the diagnosis result in a knowledge base for incremental fine-tuning of subsequent models.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine blade fault diagnosis technology, specifically relating to a wind turbine blade fault diagnosis method based on voiceprint recognition. Background Technology

[0002] As the core component of wind turbines for capturing wind energy, wind turbine blades endure multiple challenges such as alternating loads, dust erosion, and lightning strikes over long periods. Their health directly determines the operational safety and power generation efficiency of the turbine. Among related technologies, blade health monitoring has formed a collaborative system integrating contact sensing, image visual inspection, and traditional acoustic detection. The non-contact monitoring method based on blade aerodynamic noise collects acoustic signatures using microphones and combines these with artificially designed features such as Mel-frequency cepstral coefficients and spectral centroids for damage assessment, covering the entire process from signal acquisition to feature extraction.

[0003] However, existing acoustic monitoring methods directly employ fixed acoustic features and shallow classification models without constructing a multi-channel collaborative sensing and rotation cycle synchronous analysis mechanism. This results in limited feature extraction capabilities, making it difficult to distinguish complex damage modes such as cracking, icing, and lightning ablation. At the same time, the environmental noise in wind farms is complex, and traditional filtering methods cannot effectively separate the target signal from the background noise. Furthermore, the generalization ability across operating conditions is weak, and the acoustic pattern distribution varies significantly between different units and wind farms. Traditional models are difficult to transfer and apply, ultimately affecting the accurate identification of early blade damage and the efficiency of operation and maintenance decisions. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose a method for diagnosing wind turbine blade faults based on voiceprint recognition.

[0006] The second objective of this invention is to propose a wind turbine blade fault diagnosis device based on voiceprint recognition.

[0007] The third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention proposes a method for diagnosing wind turbine blade faults based on voiceprint recognition, comprising:

[0010] S1. Sound acquisition devices are deployed at multiple key locations of the wind turbine to synchronously acquire blade rotation azimuth angle signals. The acquired raw sound signals are synchronously trimmed and framed using the blade rotation cycle as a unit to obtain framed sound signals. S2, perform adaptive noise suppression and rotation synchronization enhancement processing on the framed audio signal, extract multi-level voiceprint features, and encode the voiceprint features into a unified multimodal feature vector; S3, input the multimodal feature vector into the domain-fine-tuned large language model, and use the audio understanding and semantic reasoning capabilities of the large language model to identify the type and severity of leaf damage; S4 outputs a structured diagnostic result that includes damage type, confidence level, severity rating, interpretability description, and maintenance recommendations. The diagnostic result is stored in a knowledge base for subsequent incremental fine-tuning of the model.

[0011] In one embodiment of the present invention, the plurality of preset positions include at least three of the following: the bottom of the tower, the middle of the tower, the outside of the nacelle, and the tower foundation; the sound acquisition device includes a microphone, with a sampling rate not lower than a preset threshold and a signal-to-noise ratio not lower than a preset threshold; the blade rotation azimuth angle signal is synchronously acquired by a blade azimuth angle encoder.

[0012] In one embodiment of the present invention, the adaptive noise suppression and rotation synchronization enhancement processing includes: Beamforming algorithms are used to spatially filter multi-channel audio signals to enhance the target direction signal. Variational mode decomposition is used to decompose the signal into multiple intrinsic mode functions, separating aerodynamic noise, mechanical noise and environmental noise components; The signal-to-noise ratio is improved by using the blade azimuth angle signal to perform time-domain synchronous averaging of the periodic signal.

[0013] In one embodiment of the present invention, the extraction of multi-level voiceprint features includes: Extract at least one traditional acoustic feature from the Mel frequency cepstral coefficients and their first-order and second-order differences, spectral centroid, spectral roll-off, spectral flux, and subband energy ratio. The Mel language spectrogram was converted into an image format, preserving the three-dimensional information of time-frequency-energy. Traditional acoustic features are combined with time-frequency image features to form a unified multimodal feature vector.

[0014] In one embodiment of the present invention, inputting the multimodal feature vector into a domain-fine-tuned large language model includes: The large language model is either a multimodal model with audio understanding capabilities or a text-based large language model that receives text sequence input. The fine-tuning of the domain employs low-rank adaptive technology, injecting trainable adaptation modules while keeping the base parameters unchanged; the fine-tuning dataset includes normal blade operation sound samples, typical blade damage sound samples, and environmental interference sound samples; the typical blade damage includes at least one of surface cracks, icing, lightning ablation, trailing edge cracking, and gel coat peeling.

[0015] In one embodiment of the present invention, the structured diagnostic results also include a diagnostic basis description and a suggested maintenance window period; the knowledge base is used to store each diagnostic result and subsequent actual inspection and verification data, supporting incremental fine-tuning of the large language model adaptation layer and continuous optimization of the damage-soundprint mapping relationship; the diagnostic results are pushed to the wind farm monitoring system through the API interface to trigger UAV re-inspection, generate maintenance work orders, or update blade health records.

[0016] To achieve the above objectives, a second aspect of the present invention provides a wind turbine blade fault diagnosis device based on voiceprint recognition, comprising: The sound acquisition and preprocessing module is used to deploy sound acquisition devices at multiple key locations of the wind turbine to synchronously acquire blade rotation azimuth angle signals, and to synchronously trim and frame the acquired raw sound signals in units of blade rotation cycles to obtain framed sound signals. The feature extraction and encoding module is used to perform adaptive noise suppression and rotation synchronization enhancement processing on the framed audio signal, extract multi-level voiceprint features, and encode the voiceprint features into a unified multimodal feature vector; The large language model inference module is used to input the multimodal feature vector into the domain-fine-tuned large language model, and use the audio understanding and semantic reasoning capabilities of the large language model to identify the type and severity of leaf damage. The results output and update module is used to output structured diagnostic results that include damage type, confidence level, severity level, interpretability description and operation and maintenance suggestions, and store the diagnostic results in the knowledge base for subsequent incremental fine-tuning of the model.

[0017] A method and apparatus for diagnosing wind turbine blade faults based on voiceprint recognition according to an embodiment of the present invention has the following beneficial effects: Non-contact, low-cost deployment: No sensors need to be installed on the blades; microphones can be installed using existing tower or nacelle locations. This makes retrofitting existing units convenient and cost-effective.

[0018] High recognition accuracy: LLM's powerful multimodal understanding capabilities can extract deep features from complex voiceprints, resulting in higher recognition accuracy and lower false alarm rate compared to traditional machine learning models.

[0019] Fine-grained damage type identification: It can not only determine whether there is an anomaly, but also identify the specific damage type (cracks / icing / lightning strikes, etc.), providing a basis for precise operation and maintenance.

[0020] Semantic understanding and interpretability: LLM can output diagnostic descriptions and maintenance suggestions in natural language, and maintenance personnel can understand the diagnostic conclusions without professional acoustic knowledge.

[0021] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, for implementing a wind turbine blade fault diagnosis method based on voiceprint recognition as described in the first aspect embodiment.

[0022] To achieve the above objectives, the fourth aspect of this application proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a wind turbine blade fault diagnosis method based on voiceprint recognition as described in the first aspect embodiment.

[0023] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a wind turbine blade fault diagnosis method based on voiceprint recognition according to an embodiment of the present invention; Figure 2 This is a structural diagram of a wind turbine blade fault diagnosis device based on voiceprint recognition according to an embodiment of the present invention; Figure 3 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] The following description, with reference to the accompanying drawings, describes a method and apparatus for diagnosing wind turbine blade faults based on voiceprint recognition, according to an embodiment of the present invention.

[0028] Example 1 Figure 1 This is a flowchart of a wind turbine blade fault diagnosis method based on voiceprint recognition according to an embodiment of the present invention, as shown below. Figure 1 As shown, it includes: S1. Sound acquisition devices are deployed at multiple key locations of the wind turbine to synchronously acquire blade rotation azimuth angle signals. The acquired raw sound signals are then synchronously trimmed and framed using the blade rotation cycle as a unit to obtain framed sound signals.

[0029] In some implementations, during the fault diagnosis of wind turbine blades, it is first necessary to construct a multi-location collaborative acoustic sensing network to obtain raw acoustic information that can comprehensively reflect the blade's operating status and potential damage characteristics. This method deploys sound acquisition devices at multiple key spatial locations of the wind turbine. These locations are selectively set in areas that can effectively capture aerodynamic noise, structural conducted sound, and near-field acoustic signals generated by blade rotation, thereby forming an acquisition array covering different sound source propagation paths.

[0030] Simultaneously, the system acquires the rotation azimuth angle signal of the blades, which is used to establish a precise time correspondence between the acoustic signal and the mechanical motion cycle of the blades. Based on the acquired azimuth angle information, the acquired raw sound signal is synchronously trimmed and framed, using the blade rotation cycle as the basic time unit. This divides the continuous sound stream into segmented signal units with clear physical meaning, synchronized with the blade rotation motion. This processing ensures that each framed sound signal segment corresponds to the acoustic response of the blade at a specific rotation phase, providing a structured data foundation for subsequent noise suppression, feature extraction, and rotation cycle-based synchronous enhancement.

[0031] As one implementation method, the sound acquisition device can be deployed at the bottom of the tower, the middle of the tower, the outside of the nacelle, and the tower foundation, and equipped with a windproof cover and a high dynamic range microphone. At the same time, the blade azimuth encoder is used to achieve precise alignment of the acoustic signal and the rotation cycle, thereby completing the synchronous trimming and framing of the original sound signal.

[0032] S2, perform adaptive noise suppression and rotation synchronization enhancement processing on the framed audio signal, extract multi-level voiceprint features, and encode the voiceprint features into a unified multimodal feature vector.

[0033] The framed audio signal undergoes adaptive noise suppression and rotation synchronization enhancement processing to effectively separate the target acoustic signal related to the blade's operating state from complex environmental noise and improve its signal-to-noise ratio and the discernibility of its periodic characteristics. Adaptive noise suppression utilizes the spatial correlation between multi-channel signals and the statistical characteristics of the noise field. Through spatial filtering or adaptive filtering algorithms, it suppresses background noise interference from non-target directions while preserving the integrity of the signal in the target direction, thereby obtaining a purer acoustic signal. Rotation synchronization enhancement uses synchronously acquired blade azimuth angle signals, with the blade rotation period as a reference, to perform time-domain synchronous averaging or phase-aligned accumulation on the noise-suppressed signal. This enhances the periodic acoustic components with a defined phase relationship to the blade rotation period, further suppressing non-periodic random noise.

[0034] Based on this, multi-level acoustic signature features are extracted from the processed signal. These features encompass multiple dimensions, including the time domain, frequency domain, time-frequency domain, and perceptual domain. They include, but are not limited to, cepstral coefficients reflecting the resonance characteristics of the acoustic duct, spectral features describing the distribution and dynamic changes of spectral energy, sub-band energy distribution features based on the auditory perception model, and time-frequency image features that retain three-dimensional information of time, frequency, and energy. These features characterize the essential properties of the acoustic signals of blade operation from different perspectives, effectively representing the differences in acoustic signatures under normal operating conditions and various damage states. Finally, the extracted multi-level acoustic signature features are dimensionally aligned and numerically normalized, and encoded into a unified multimodal feature vector through feature splicing or attention fusion mechanisms. This feature vector serves as the input embedding for subsequent large-scale language models to perform acoustic signature decomposition and damage identification.

[0035] As one implementation, the multi-level voiceprint features may include Mel frequency cepstral coefficients and their differences, spectral centroid, spectral roll-off, spectral flux, subband energy ratio based on auditory perception wavelet packet decomposition, and Mel speech spectrogram. After converting the Mel speech spectrogram into an image format, the dynamic coding difference features of adjacent periodic pixels are extracted, and finally the above features are concatenated into a unified multimodal feature vector.

[0036] S3. Input the multimodal feature vector into the domain-fine-tuned large language model, and use the audio understanding and semantic reasoning capabilities of the large language model to identify the type and severity of leaf damage.

[0037] In some implementations, the multimodal feature vectors are input into a domain-fine-tuned large language model, which possesses the ability to understand audio information and reason about complex semantic relationships. The core function of this large language model is its ability to receive and parse the structured or semi-structured multimodal feature vectors generated by the preceding steps. These feature vectors integrate multi-dimensional information such as the time domain, frequency domain, and energy distribution of the sound signal. Internally, through its vast parameter space and the general knowledge learned during the pre-training phase, the model can perform deep pattern matching and logical association analysis on these input features.

[0038] Specifically, this model does not simply map features to labels. Instead, based on a semantic understanding of the physical mechanisms behind acoustic phenomena, it associates specific patterns in the input feature vector with predefined or fine-tuned acoustic representations of blade damage. Through domain fine-tuning, the model is endowed with expertise specific to the operating scenarios of wind turbine blades, enabling it to distinguish between acoustic fluctuations under normal operating conditions and specific acoustic feature changes caused by different damage types such as cracks, icing, and lightning strikes. Utilizing its reasoning capabilities, the model comprehensively evaluates the contribution and confidence of each component in the feature vector, thereby achieving an accurate assessment of the current health status of the blade and outputting identification results regarding the type and severity of damage.

[0039] As one implementation method, the large language model can be a model based on an audio-text multimodal architecture, directly receiving image features such as spectrograms; or, it can be a large language model that encodes the extracted voiceprint features into a text sequence and inputs it into the text fine-tuning model, using fine-tuning techniques such as low-rank adaptive (LoRA) to accurately identify various typical blade damage types and their corresponding severity levels, including surface cracks, icing, lightning ablation, trailing edge cracks.

[0040] S4 outputs a structured diagnostic result that includes damage type, confidence level, severity rating, interpretability description, and maintenance recommendations. The diagnostic result is stored in a knowledge base for subsequent incremental fine-tuning of the model.

[0041] After obtaining the recognition results generated by the large language model based on voiceprint feature analysis, the method further performs diagnostic result output and knowledge management steps. Specifically, the method structures the raw inference results output by the large language model to generate a diagnostic report containing multi-dimensional information. This report at least covers the identified leaf damage type, the confidence assessment corresponding to the damage type, the severity level determined based on preset grading standards, the interpretability explanation of the acoustic features or reasoning logic used to interpret the diagnostic conclusions, and the recommended operation and maintenance suggestions or maintenance window period for the current damage status. This structured diagnostic result is encapsulated in a standardized data format to facilitate subsequent interaction and parsing between systems.

[0042] Simultaneously, this method stores each generated diagnostic result, along with its corresponding original acoustic signature feature data or feature vector, and any subsequent potential verification results, into a continuously evolving knowledge base. This knowledge base serves as the core data source for model iterative optimization. The historical diagnostic data and verification feedback stored in this knowledge base are used for incremental fine-tuning of the large language model. Specifically, newly accumulated sample data is used to update the model's adaptation layer or specific parameters, thereby continuously optimizing the model's mapping relationship to the acoustic signature features of blade damage. This enhances the model's ability to identify novel damage patterns and improves its generalization performance across different operating conditions or different units, forming a closed-loop evolution mechanism from diagnostic output to knowledge accumulation and then to model optimization.

[0043] In one implementation, the structured diagnostic results may include damage type and confidence level, severity classification of four levels of warning (green / yellow / orange / red), interpretability description based on acoustic features, and specific operation and maintenance steps. Each diagnostic result is stored in a knowledge base for subsequent incremental fine-tuning of the model adaptation layer.

[0044] An embodiment of the present invention provides a wind turbine blade fault diagnosis method based on voiceprint recognition, which can realize non-contact, low-cost blade damage monitoring, improve the fineness and accuracy of damage type identification, and output interpretable diagnostic results and operation and maintenance suggestions.

[0045] Example 2 The following describes in detail, with reference to the accompanying drawings, a wind turbine blade fault diagnosis method based on voiceprint recognition according to an embodiment of the present invention. The technical solution proposed by the present invention mainly includes the following steps: Step 1: Multi-location sound acquisition and preprocessing.

[0046] Deploying sound acquisition equipment arrays at key locations of wind turbine units, specifically including: Bottom of the tower: Collect low-frequency aerodynamic noise and structurally conducted sound.

[0047] Middle section of the tower (1 / 3 to 2 / 3 of the height from the ground): Main data acquisition channel, taking into account both signal-to-noise ratio and wind noise resistance.

[0048] Outside the nacelle: Collect near-field aerodynamic sounds as the blades rotate through the nacelle.

[0049] Reference microphone at the tower foundation: used for environmental noise phase cancellation reference.

[0050] Each acquisition location is equipped with a windproof cover and a high dynamic range microphone (sampling rate ≥44.1kHz, signal-to-noise ratio ≥70dB). Simultaneous acquisition of blade azimuth encoder signals ensures alignment of the acoustic signals with the blade rotation period.

[0051] The acquired raw sound signals undergo the following preprocessing steps: Synchronous cropping and framing: Signal segmentation is performed using the blade rotation period (1 / P, where P is the number of blades) as the unit.

[0052] Adaptive noise suppression: Beamforming algorithms are used to spatially filter multi-channel signals to enhance the target direction signal.

[0053] Variational mode decomposition: decomposes the signal into several intrinsic mode functions, separating aerodynamic noise, mechanical noise and environmental noise components.

[0054] Rotation synchronization enhancement: The periodic signal is synchronously averaged in the time domain using the blade azimuth angle signal to improve the signal-to-noise ratio.

[0055] Step 2: Voiceprint feature extraction and encoding.

[0056] Extracting multi-level voiceprint features from the preprocessed audio signal: Traditional acoustic feature extraction (as a supplementary feature to LLM): Mel frequency cepstral coefficients (MFCC): 13-dimensional static coefficients + first / second order difference; spectral centroid, spectral roll-off, spectral flux; subband energy ratio (using auditory perception wavelet packet decomposition to divide into 24 auditory subbands); Mel spectrogram: retaining three-dimensional information of time-frequency-energy.

[0057] Time-frequency image coding: converts the Mel speech spectrogram into an image format, preserving the dynamic coding difference features of adjacent periodic pixels, which is used to capture the periodic changes of damaged sounds.

[0058] Multimodal feature fusion: concatenates the above features into a unified feature vector F=[f MFCC ,f spec ,f energy ,f img [Embedded as audio input for LLM]

[0059] Step 3: LLM-based acoustic texture solution and damage identification.

[0060] Large language model selection and adaptation.

[0061] This method employs a large-scale language model with audio understanding capabilities as its core inference engine, including but not limited to: Audio-text multimodal LLM: Models based on architectures such as Qwen-Audio and SALMONN can directly receive raw audio or spectrograms as input.

[0062] Text-based fine-tuned LLM: The features extracted in step two are encoded into text sequences and then input into an LLM (such as the GPT series, Llama series, locally deployed ChatGLM, etc.).

[0063] To adapt to the vertical field of wind turbine blade diagnostics, a low-rank adaptive (LoRA) fine-tuning technique is adopted. While keeping the LLM base parameters unchanged, a trainable adaptation module is injected. The fine-tuning dataset includes: normal blade operating sound samples (under different wind speeds and power conditions); typical blade damage sound samples (cracks, icing, lightning strike ablation, trailing edge cracking, gelcoat peeling, etc.); and environmental interference sound samples (rain sounds, thunder, birdsong, mechanical noises, etc.).

[0064] The LLM, after fine-tuning, enables voiceprint recognition of the following typical blade damage, as shown in Table 1 below: Table 1

[0065] Number of sound acquisition devices: preferably 4-6, of which the main acquisition point in the middle of the tower adopts a microphone array (no less than 3 microphones forming a linear array) to achieve beamforming directional enhancement.

[0066] LLM deployment method: adopts an edge-cloud collaborative architecture. Real-time diagnosis is completed by a lightweight LLM at the edge (such as the quantized ChatGLM3-6B). Complex cases are uploaded to the cloud for full LLM for secondary confirmation.

[0067] Data privacy protection: The original audio data is not uploaded to the cloud; only the anonymized feature vector and diagnostic request are uploaded, which complies with industrial data security requirements.

[0068] Diagnostic cycle: A diagnostic result is output every 10 rotation cycles (approximately 30-60 seconds), achieving near real-time monitoring.

[0069] Transfer learning across operating conditions.

[0070] To address the differences in acoustic signature distribution among different generator units and wind farms, a migration diagnostic mechanism is employed. The voiceprint features of the source domain (labeled dataset) and the target domain (unit to be diagnosed) are embedded into the Grassmann manifold space; Calculate the transferability between the feature vectors of the source domain and the target domain; A transfer objective function is constructed based on the maximum mean difference to adapt to the feature distribution; By fine-tuning the LLM adaptation layer using the migrated features, cross-unit diagnostic capabilities can be achieved.

[0071] Step 4: Outputting diagnostic results and updating the knowledge base.

[0072] Multimodal results are presented.

[0073] LLM outputs structured diagnostic results, including: damage type and confidence level; severity level (green / yellow / orange / red four-level warning); interpretability explanation (based on which acoustic features the conclusions were drawn); and maintenance recommendations (specific operating steps and suggested maintenance windows).

[0074] Closed-loop update of the knowledge base.

[0075] Each diagnostic result (including subsequent actual examination and verification) is stored in a knowledge base for: incremental fine-tuning of the LLM adaptation layer; continuous optimization of the damage-voiceprint mapping relationship; and accumulation and identification of novel damage patterns.

[0076] It works in conjunction with the intelligent operation and maintenance system.

[0077] The diagnostic results are pushed to the wind farm SCADA system or intelligent inspection system via API interface, which automatically triggers drone re-inspection (when there is a high confidence anomaly); generates maintenance work orders and assigns them to the operation and maintenance team; and links the blade digital passport to update the health record.

[0078] Example 3 To achieve the above embodiments, such as Figure 2 As shown, this embodiment also provides a wind turbine blade fault diagnosis device 10 based on voiceprint recognition. The device 10 includes a sound acquisition and preprocessing module 100, a feature extraction and encoding module 200, a large language model inference module 300, and a result output and update module 400.

[0079] The sound acquisition and preprocessing module 100 is used to deploy sound acquisition devices at multiple key locations of the wind turbine, synchronously acquire blade rotation azimuth angle signals, and synchronously trim and frame the acquired raw sound signals in units of blade rotation cycle to obtain framed sound signals. The feature extraction and encoding module 200 is used to perform adaptive noise suppression and rotation synchronization enhancement processing on the framed audio signal, extract multi-level voiceprint features, and encode the voiceprint features into a unified multimodal feature vector. The large language model reasoning module 300 is used to input the multimodal feature vector into the domain-fine-tuned large language model, and use the audio understanding and semantic reasoning capabilities of the large language model to identify the type and severity of leaf damage. The results output and update module 400 is used to output structured diagnostic results including damage type, confidence level, severity level, interpretability description and operation and maintenance suggestions, and store the diagnostic results in the knowledge base for subsequent incremental fine-tuning of the model.

[0080] This invention discloses a wind turbine blade fault diagnosis device based on voiceprint recognition, which can realize non-contact, low-cost blade damage monitoring, improve the fineness and accuracy of damage type identification, and output interpretable diagnostic results and maintenance suggestions.

[0081] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 3 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.

[0082] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0083] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0084] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for diagnosing wind turbine blade faults based on voiceprint recognition, characterized in that, include: S1. Sound acquisition devices are deployed at multiple key locations of the wind turbine to synchronously acquire blade rotation azimuth angle signals. The acquired raw sound signals are synchronously trimmed and framed using the blade rotation cycle as a unit to obtain framed sound signals. S2, perform adaptive noise suppression and rotation synchronization enhancement processing on the framed audio signal, extract multi-level voiceprint features, and encode the voiceprint features into a unified multimodal feature vector; S3, input the multimodal feature vector into the domain-fine-tuned large language model, and use the audio understanding and semantic reasoning capabilities of the large language model to identify the type and severity of leaf damage; S4 outputs a structured diagnostic result that includes damage type, confidence level, severity rating, interpretability description, and maintenance recommendations. The diagnostic result is stored in a knowledge base for subsequent incremental fine-tuning of the model.

2. The method according to claim 1, characterized in that, The multiple key locations include at least three of the following: the bottom of the tower, the middle of the tower, the outside of the nacelle, and the tower foundation; the sound acquisition device includes a microphone, with a sampling rate not lower than a preset threshold and a signal-to-noise ratio not lower than a preset threshold; the blade rotation azimuth angle signal is synchronously acquired through a blade azimuth angle encoder.

3. The method according to claim 1, characterized in that, The adaptive noise suppression and rotation synchronization enhancement processing includes: Beamforming algorithms are used to spatially filter multi-channel audio signals to enhance the target direction signal. Variational mode decomposition is used to decompose the signal into multiple intrinsic mode functions, separating aerodynamic noise, mechanical noise and environmental noise components; The signal-to-noise ratio is improved by using the blade azimuth angle signal to perform time-domain synchronous averaging of the periodic signal.

4. The method according to claim 1, characterized in that, The extraction of multi-level voiceprint features includes: Extract at least one traditional acoustic feature from the Mel frequency cepstral coefficients and their first-order and second-order differences, spectral centroid, spectral roll-off, spectral flux, and subband energy ratio. The Mel language spectrogram was converted into an image format, preserving the three-dimensional information of time-frequency-energy. Traditional acoustic features are combined with time-frequency image features to form a unified multimodal feature vector.

5. The method according to claim 1, characterized in that, The step of inputting the multimodal feature vector into a domain-fine-tuned large language model includes: The large language model is either a multimodal model with audio understanding capabilities or a text-based large language model that receives text sequence input. The fine-tuning of the domain employs low-rank adaptive technology, injecting trainable adaptation modules while keeping the base parameters unchanged; the fine-tuning dataset includes normal blade operation sound samples, typical blade damage sound samples, and environmental interference sound samples; the typical blade damage includes at least one of surface cracks, icing, lightning ablation, trailing edge cracking, and gel coat peeling.

6. The method according to claim 1, characterized in that, The structured diagnostic results also include explanations of the diagnostic basis and suggested maintenance windows; the knowledge base is used to store each diagnostic result and subsequent actual inspection and verification data, supporting incremental fine-tuning of the large language model adaptation layer and continuous optimization of the damage-voiceprint mapping relationship; The diagnostic results are pushed to the wind farm monitoring system via API interface to trigger drone re-inspection, generate maintenance work orders, or update blade health records.

7. A wind turbine blade fault diagnosis device based on voiceprint recognition, characterized in that, include: The sound acquisition and preprocessing module is used to deploy sound acquisition devices at multiple key locations of the wind turbine to synchronously acquire blade rotation azimuth angle signals, and to synchronously trim and frame the acquired raw sound signals in units of blade rotation cycles to obtain framed sound signals. The feature extraction and encoding module is used to perform adaptive noise suppression and rotation synchronization enhancement processing on the framed audio signal, extract multi-level voiceprint features, and encode the voiceprint features into a unified multimodal feature vector; The large language model inference module is used to input the multimodal feature vector into the domain-fine-tuned large language model, and use the audio understanding and semantic reasoning capabilities of the large language model to identify the type and severity of leaf damage. The results output and update module is used to output structured diagnostic results that include damage type, confidence level, severity level, interpretability description and operation and maintenance suggestions, and store the diagnostic results in the knowledge base for subsequent incremental fine-tuning of the model.

8. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement a wind turbine blade fault diagnosis method based on voiceprint recognition as described in any one of claims 1-6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a wind turbine blade fault diagnosis method based on voiceprint recognition as described in any one of claims 1-6.

Images