Wind turbine generator voiceprint fault recognition method

Abstract

The invention provides a wind turbine generator voiceprint fault recognition method, and relates to the technical field of wind turbine generator state monitoring and fault diagnosis, and the method comprises the steps: carrying out the noise reduction of an original audio signal through variational mode decomposition, screening a target mode of which the frequency, energy and kurtosis accord with features, and reconstructing the signal; extracting a Mel frequency cepstrum coefficient and a sensing noise robust coefficient, and generating multi-dimensional voiceprint data in combination with statistical characteristics such as a frequency spectrum gravity center, a spectrum entropy, energy, kurtosis and a zero-crossing rate; constructing a support set based on the prototype network, realizing small sample fault classification by calculating the Euclidean distance between the feature vector and the prototype vector, and outputting a preliminary result; judging whether the voiceprint is abnormal according to a preset threshold value, if so, storing the voiceprint into a dynamic abnormal voiceprint knowledge base; frequently occurring abnormal samples are manually labeled and added into a support set, the prototype network is retrained to update the model, and continuous optimization of the fault recognition capability is achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wind turbine condition monitoring and fault diagnosis, and particularly relates to a wind turbine voiceprint fault recognition method. BACKGROUND

[0002] As an important part of clean energy, the running stability of a wind turbine is directly related to power generation efficiency and operation and maintenance cost. Traditional condition monitoring mainly relies on SCADA systems and vibration monitoring systems (CMS) to collect structured parameters such as temperature, vibration and current to perform fault early warning. However, these methods have low sensitivity to early weak faults (such as slight bearing wear and slight gear crack), and it is difficult to achieve early warning.

[0003] In recent years, voiceprint recognition technology has been gradually applied to equipment health monitoring due to its non-contact and rich information. However, the running environment of a wind turbine is complex, and a large amount of wind noise, electromagnetic interference and background noise are mixed in the audio signal, making it difficult to extract acoustic features. In addition, the cost of obtaining fault samples is high and the labeling period is long in actual operation and maintenance, and most fault types only have a small amount of available samples. Traditional deep learning models (such as CNN and RNN) rely on large-scale labeled data, and it is difficult to effectively model in a "small sample" scenario.

[0004] In the prior art, some methods attempt to extract audio features using traditional signal processing methods (such as wavelet transform and short-time Fourier transform) and combine support vector machines (SVM) for classification. However, this type of method has poor feature robustness in a strong noise environment and cannot adapt to the dynamic recognition needs of new fault types. Although small sample learning methods such as prototype networks have made progress in image recognition, their application in wind turbine voiceprint recognition still faces problems such as insufficient feature extraction, weak noise suppression ability and lack of model updating mechanism. SUMMARY

[0005] The application provides a wind turbine voiceprint fault recognition method to solve one of the above technical problems.

[0006] The technical solution adopted by the application is as follows: The application provides a wind turbine voiceprint fault recognition method, which comprises: An original audio signal of a wind turbine in operation is collected, and the original audio signal is decomposed into a plurality of intrinsic modal components through variational modal decomposition. Target modes are obtained by screening the intrinsic modal components according to the frequency distribution, energy density and kurtosis of the modes, so as to reconstruct a denoised audio signal; Mel frequency cepstral coefficient calculation is performed on the denoised audio signal, and a perception noise robust coefficient algorithm is combined to enhance feature robustness. At the same time, spectral barycenter, spectral entropy, energy, kurtosis and zero-crossing rate statistical features are extracted to generate multi-dimensional voiceprint data. Based on the prototype network, the voiceprint feature vectors of known fault types are abstracted into prototype vectors to form a support set. The extracted multidimensional voiceprint features are converted into feature vectors. By calculating the Euclidean distance between the feature vectors and each prototype vector, the similarity between the feature vectors and known fault categories is determined. The corresponding fault category is matched according to the minimum distance value, and the preliminary classification result is output. Based on the preliminary classification results, a preset threshold is used to determine whether the feature vector belongs to the abnormal category; if it is marked as abnormal, the abnormal multidimensional voiceprint data is stored in the dynamically updated abnormal voiceprint knowledge base. By manually labeling frequently occurring anomalous samples and adding them to the support set, the prototype network can be retrained or adjusted to update the prototype vector.

[0007] According to one embodiment of this application, in the step of decomposing the original audio signal into multiple intrinsic mode components by variational mode decomposition, the number of intrinsic mode components is set to 3-7, and the target mode is selected according to the following conditions: The intrinsic mode components that match the vibration frequency of key components of the wind turbine were selected, with the center frequency range being 50Hz-2000Hz. Retain intrinsic mode components with energy density higher than a preset threshold, and remove low-energy noise components; Intrinsic mode components are screened using kurtosis indices, retaining those with kurtosis values ​​greater than 1.5 to preserve non-Gaussian features in the signal.

[0008] According to one embodiment of this application, in the step of calculating the Mel frequency cepstral coefficients of the denoised audio signal, the calculation parameters of the Mel frequency cepstral coefficients include: The audio signal is divided into time-domain segments with a frame length of 20ms and a frame shift of 10ms; It employs 24-40 Mel filters, covering a frequency range of 20Hz-4000Hz; After taking the logarithm of the power spectrum, a discrete cosine transform is performed to extract the 12-20 dimension Mel frequency cepstral coefficients, and the first 12 dimensions are retained as the main features.

[0009] According to one embodiment of this application, in the step of enhancing feature robustness by combining the perceptual noise robustness coefficient algorithm, the implementation of the perceptual noise robustness coefficient includes: The power spectral density of noise is estimated from environmental noise samples and separated from the spectrum of the audio signal; The perceived noise robustness coefficients are weighted, and the weighting coefficients are correlated with the signal-to-noise ratio of the noise. The compensation parameters are dynamically adjusted based on real-time noise changes to ensure that the perceived noise robustness coefficient remains stable in complex noise environments.

[0010] According to one embodiment of this application, in the step of extracting the statistical features of spectral centroid, spectral entropy, energy, kurtosis, and zero-crossing rate, the calculation method for each feature is as follows: The formula is calculated by weighted averaging of spectral energy. , where f i For the frequency point, E i For the corresponding energy; The uncertainty of the spectral distribution is calculated based on the Shannon entropy formula, which characterizes the complexity of the signal; The number of times an audio signal passes through zero within a unit of time is counted to reflect the distribution characteristics of high-frequency components.

[0011] According to one embodiment of this application, in the step of abstracting the voiceprint feature vector of a known fault type into a prototype vector based on a prototype network, the method for generating the prototype vector includes: Z-score standardization is applied to the feature vectors supporting the same type of fault in the set, so that their mean is 0 and their variance is 1. The prototype vector is generated by taking the mean of the feature vectors of the same type. If the number of samples of the same type is insufficient, the K-nearest neighbor weighted average method is used to supplement it. If the eigenvector dimension is higher than 50, principal component analysis is used to reduce the dimension to 20-40 to reduce computational complexity.

[0012] According to one embodiment of this application, the preset threshold is set based on the Euclidean distance distribution under normal conditions, and is dynamically adjusted through the following steps: The 95th percentile was calculated based on the Euclidean distance from historical normal samples and used as the initial threshold. Based on the Euclidean distance distribution of real-time operational data, the threshold is updated every 24 hours, and the correction formula is as follows: ; Where μ is the current mean, σ is the standard deviation, and k is the empirical coefficient.

[0013] According to one embodiment of this application, in the step of determining similarity by calculating the Euclidean distance between the feature vector and each prototype vector, the calculation of the Euclidean distance includes: Normalize the Euclidean distance to the [0,1] interval based on the maximum and minimum values; Weights are assigned based on the importance of the feature dimensions; If the minimum distance is less than the preset threshold, it is considered a match; otherwise, it is marked as an anomaly.

[0014] A second aspect of this application provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps described in the method.

[0015] A third aspect of this application provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method as described.

[0016] Due to the adoption of the above technical solution, the beneficial effects achieved by this application are as follows: This application decomposes the original audio signal into multiple intrinsic mode components using variational mode decomposition (VMD), and selects target modes based on frequency distribution, energy density, and kurtosis, effectively eliminating high-frequency noise and low-frequency background interference. This preprocessing method retains effective acoustic components related to the vibration characteristics of key components (such as gearboxes and bearings), significantly improving the clarity of the reconstructed audio signal and providing high-quality input for subsequent feature extraction. It also solves the problem of feature extraction distortion in strong noise environments using traditional methods.

[0017] In the feature extraction stage, both Mel frequency cepstral coefficients (MFCC) and perceptual noise robustness coefficients (PNCC) algorithms are used. The former preserves speech features in frequency bands sensitive to human hearing, while the latter enhances the adaptability of features to environmental changes through noise modeling. Simultaneously, statistical features such as spectral centroid, spectral entropy, energy, kurtosis, and zero-crossing rate are introduced to characterize voiceprint changes from multiple dimensions in the time and frequency domains. This multi-dimensional feature fusion strategy significantly improves the ability of feature vectors to distinguish device operating states, maintaining stability even under complex conditions such as wind speed changes and rain / snow, overcoming the vulnerability of single features to interference.

[0018] A small-sample classification model is built based on a prototype network. The acoustic signature features of each known fault type are abstracted into prototype vectors to form a support set. Similarity matching is performed by calculating the Euclidean distance between the input feature vector and each prototype vector, requiring only a small number of samples to complete the model construction. This mechanism effectively reduces the dependence on large-scale labeled data, solves the practical problem of "faults but no data" in wind power sites, and achieves a classification accuracy of over 90% even when there are fewer than 20 samples for each fault type.

[0019] The initial classification results are judged by setting a preset threshold. When the distance between the feature vector and the nearest prototype exceeds the threshold, it is marked as an anomaly category, and the multidimensional voiceprint data is stored in a dynamically updated anomaly voiceprint knowledge base. This mechanism can identify unknown faults or novel anomalies that are not in the support set, making up for the limitation of traditional classification models that can only identify known categories, and improving the system's fault tolerance and early warning capabilities.

[0020] By manually labeling frequently occurring anomalous samples and adding them to the support set, the prototype network is retrained or adjusted to update the prototype vectors, forming a closed-loop learning process. This self-learning mechanism enables the system to "learn while running," gradually accumulating new fault knowledge and adapting to long-term maintenance needs such as equipment aging and changes in operating conditions. It avoids the drawbacks of traditional systems that require periodic shutdowns for retraining, significantly improving the intelligence and sustainability of the diagnostic system. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a wind turbine acoustic signature fault identification method provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0022] Figure label: 810, Processor; 820, Communication interface; 830, Memory; 840, Communication bus. Detailed Implementation

[0023] To more clearly illustrate the overall concept of this application, a detailed explanation is provided below with reference to the accompanying drawings.

[0024] Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application may also be implemented in other ways different from those described herein. Therefore, the scope of protection of this application is not limited to the specific embodiments disclosed below. It should be noted that, unless otherwise specified, the embodiments of this application and the features thereof can be combined with each other.

[0025] In this application, unless otherwise expressly specified and limited, the "above" or "below" of the second feature can mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. In the description of this specification, references to terms such as "an embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. 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 can be combined in any suitable manner in one or more embodiments or examples. Example

[0026] likeFigure 1 As shown, a method for identifying acoustic fingerprint faults in wind turbine generators includes: The original audio signal of the wind turbine is collected during operation and decomposed into multiple intrinsic mode components through variational mode decomposition. The intrinsic mode components are screened according to the frequency distribution, energy density and kurtosis of the modes to obtain the target mode, so as to reconstruct the noise-reduced audio signal.

[0027] As mentioned above, a high-sensitivity microphone array deployed near the wind turbine tower, nacelle, or blades is used to collect the raw audio signals of the equipment under normal operation or fault conditions. Due to the presence of non-stationary noise such as strong wind noise, electromagnetic interference, and mechanical vibration coupling at wind power sites, the raw signals usually have a low signal-to-noise ratio, and directly using them for feature extraction can lead to misjudgments.

[0028] Therefore, Variational Mode Decomposition (VMD) is used to adaptively decompose the original signal. VMD is a non-recursive signal decomposition method that can decompose non-stationary audio signals into several intrinsic mode components (IMFs) with finite bandwidth. Each IMF represents the oscillation mode of the signal within a specific frequency band and has a clear physical meaning. Compared with traditional wavelet transform or empirical mode decomposition (EMD), VMD has stronger anti-mode aliasing ability and can more accurately separate different frequency components.

[0029] After obtaining multiple IMF components, it is necessary to further filter out the "target modes" related to the equipment's operating status. The screening criteria include three core indicators: Frequency distribution: Analyze the center frequency of each IMF and retain the modes located within the typical vibration frequency range of key components of the wind turbine (such as gearbox, main bearing, generator) (e.g., 50Hz–2000Hz). Energy density: Calculate the energy proportion of each IMF, prioritize retaining modes with high energy concentration and large contribution to the overall signal, and remove noise components with weak energy. Kurtosis: As a measure of the non-Gaussianity of a signal, high kurtosis values ​​usually correspond to the impact component and are closely related to the transient acoustic response caused by early faults (such as pitting and cracks). Therefore, the IMF with kurtosis is significantly higher than that of background noise.

[0030] Finally, the selected target modes are linearly superimposed to reconstruct a denoised audio signal. This signal retains the essential acoustic characteristics of the device's operation while effectively suppressing environmental noise and interference, providing high-quality input for subsequent feature extraction.

[0031] For example, consider the acoustic signature collection of an onshore wind turbine operating at moderate wind speeds (8–10 m / s): A 30-second raw audio signal was acquired inside the cabin using a directional microphone with a sampling frequency of 48 kHz. This signal contains noticeable wind noise (wideband random noise), gear meshing sounds (periodic pulses), bearing rotation sounds, and background electromagnetic interference.

[0032] The signal was decomposed using VMD, with a decomposition level of 5, resulting in 5 IMF components. Analysis revealed that: IMF1: Center frequency approximately 3500Hz, low energy, kurtosis 1.2, exhibiting high-frequency wind noise, classified as a noise mode; IMF2: Center frequency approximately 1800Hz, energy percentage 18%, kurtosis 2.1, corresponding to high-speed shaft gear meshing signal; IMF3: Center frequency approximately 650Hz, energy percentage 32%, kurtosis 2.8, matching the main bearing rotation frequency; IMF4: Center frequency approximately 120Hz, energy percentage 25%, kurtosis 1.6, corresponding to low-frequency vibration of a generator; IMF5: Center frequency approximately 30Hz, energy percentage 5%, kurtosis 1.1, representing a low-frequency background disturbance.

[0033] Based on a comprehensive analysis of frequency distribution, energy density, and kurtosis, IMF2, IMF3, and IMF4 were selected as target modes, while the other two IMFs were discarded. Reconstructing these three components yielded a clear audio signal with a signal-to-noise ratio improvement exceeding 15 dB. The gear meshing and bearing acoustic signatures were significantly enhanced, making it suitable for subsequent feature extraction and fault identification.

[0034] It should be noted that, in specific implementation scenarios, the number of decomposition layers and penalty factor of VMD can be dynamically adjusted based on different machine models or operating conditions (such as low wind speed, high wind speed, and start-up / shutdown processes) to improve the adaptability of mode decomposition. For example, the number of decomposition layers can be increased under high wind speeds to better separate broadband noise.

[0035] In specific implementation scenarios, based on the above scheme, the soundprint signals collected by multiple microphones can be combined, and spatial filtering can be performed first using beamforming or blind source separation technology, followed by VMD decomposition, to further improve the ability to extract the target sound source.

[0036] In specific implementation scenarios, in addition to frequency, energy, and kurtosis, the above-mentioned schemes can also introduce indicators such as sample entropy, correlation coefficient, and Hilbert marginal spectrum as auxiliary screening conditions to improve the accuracy of target modality recognition.

[0037] In specific implementation scenarios, a lightweight VMD algorithm can be deployed on edge computing devices based on the above solutions to achieve real-time decomposition and reconstruction of audio signals, meeting the low latency requirements of wind turbine online monitoring.

[0038] In specific implementation scenarios, VMD can be cascaded with wavelet packet decomposition, empirical wavelet transform (EWT), and other methods in specific scenarios to first perform coarse decomposition and then fine screening, thereby further improving the denoising effect.

[0039] In specific implementation scenarios, based on the above scheme, for some IMFs that are not included in the target mode but have abnormal characteristics (such as sudden high energy or extremely high kurtosis), they can be marked separately and an anomaly warning can be triggered as an indirect clue to early failure.

[0040] Mel frequency cepstral coefficients are calculated for the denoised audio signal, and the feature robustness is enhanced by combining the perceptual noise robustness coefficient algorithm. At the same time, statistical features such as spectral centroid, spectral entropy, energy, kurtosis and zero-crossing rate are extracted to generate multidimensional voiceprint data.

[0041] As mentioned above, due to the complex environment at wind power sites, even after VMD noise reduction, the signal may still be affected by residual noise, operating condition fluctuations, and other factors. Therefore, it is necessary to use a variety of complementary feature extraction methods to construct a comprehensive and stable feature space.

[0042] First, Mel frequency cepstral coefficients (MFCC) are calculated. MFCC simulates the nonlinear perception of sound frequencies by the human ear, mapping linear frequencies to the Mel scale, which better reflects the sensitivity of the human auditory system to the mid-to-low frequency range. This feature can effectively capture the periodic acoustic patterns of mechanical components such as gear meshing and bearing rotation, and is particularly suitable for characterizing steady-state acoustic signature changes caused by structural vibrations.

[0043] To further enhance the stability of features in noisy environments, the Perceived Noise Robustness Coefficient (PNCC) algorithm is introduced as an enhancement method. Building upon MFCC, PNCC models and compensates for background noise by simulating the gain control mechanism of the auditory system, significantly improving the feature's resistance to non-stationary interference such as wind noise and rain noise. This algorithm is particularly suitable for situations where wind speed variations cause signal-to-noise ratio fluctuations in wind power scenarios, ensuring feature consistency under different environmental conditions.

[0044] In addition, to comprehensively characterize the dynamic characteristics of the voiceprint signal, a set of time-domain and frequency-domain statistical features were extracted simultaneously, including: Spectrum centroid: Reflects the concentrated position of signal energy on the frequency axis, and can be used to identify high-frequency energy rise caused by component wear; Spectral entropy: measures the degree of disorder in the spectrum distribution; abnormal equipment status is often accompanied by an increase in spectral entropy. Energy: Characterizes the overall change in signal strength and is sensitive to changes in load or structural loosening; Kurtosis: describes the intensity of the impact component in a signal and is highly sensitive to transient faults such as early pitting and cracks; Zero-crossing rate: the frequency at which a statistical signal crosses zero, reflecting the activity level of high-frequency components, and can be used to distinguish friction-related faults.

[0045] Finally, the MFCC, PNCC, and the aforementioned statistical features are concatenated into vectors to form a set of high-dimensional, multi-dimensional "acoustic print data," which serves as the input for the subsequent small-sample classification model. This multi-dimensional feature system takes into account both steady-state and transient, global and local acoustic characteristics, significantly improving the model's ability to identify different fault types.

[0046] For example, let's take the acoustic signature analysis of a wind turbine under normal operation and slight gear wear conditions as an example: After VMD noise reduction, a clear audio signal was obtained with a sampling rate of 48kHz, a frame length of 20ms, and a frame shift of 10ms.

[0047] Calculation of the 12-dimensional MFCC coefficients revealed that the coefficients in dimensions 3–6 increased significantly under wear conditions, indicating enhanced acoustic energy in the mid-frequency range. Simultaneously, the characteristics of PNCC were calculated, and it was found that its characteristic value fluctuation was less than 30% of that of MFCC during the process of wind speed increasing from 6 m / s to 12 m / s, which verified its stronger environmental adaptability. Further extraction of statistical features: The spectral centroid rose from 850Hz to 980Hz, indicating an increase in high-frequency components; The spectral entropy increased from 0.68 to 0.75, reflecting a more uneven spectral distribution; The kurtosis increased from 2.1 to 3.4, suggesting the presence of more impactful acoustic events; The zero-crossing rate increased from 120 times / second to 145 times / second, indicating that high-frequency vibration intensified.

[0048] The 12-dimensional MFCC, 12-dimensional PNCC, and 5-dimensional statistical features (a total of 29 dimensions) are concatenated into a feature vector to form the "multidimensional voiceprint data" of the audio frame. This vector shows significant differences between normal and fault states and can serve as effective input for classification models.

[0049] It should be noted that, in specific implementation scenarios, based on the above scheme, the feature subsets participating in the splicing can be dynamically selected according to different wind turbine models or fault types. For example, for bearing faults, wavelet packet energy entropy can be added; for friction-related faults, modulation domain features can be introduced.

[0050] In specific implementation scenarios, based on the above approach, different weights can be assigned to different features when generating multidimensional voiceprint data. For example, in high wind noise environments, the weights of PNCC and kurtosis can be increased to improve classification accuracy.

[0051] In specific implementation scenarios, based on the above solutions, audio signals can be converted into spectrogram images, from which texture features (such as gray-level co-occurrence matrices) can be extracted or deep features can be extracted through lightweight convolutional networks, and then fused with traditional handcrafted features to form a hybrid feature vector.

[0052] In specific implementation scenarios, based on the above solutions, the contribution of each feature dimension can be dynamically evaluated during system operation based on the feedback of classification results, redundant or low-discrimination features can be eliminated, and the feature space structure can be optimized.

[0053] In specific implementation scenarios, based on the above scheme, not only can single-frame features be extracted, but the feature vectors of multiple consecutive frames can also be encoded by mean, variance or LSTM to form context-aware temporal features, thereby improving the ability to identify gradual faults.

[0054] In specific implementation scenarios, based on the above scheme, before generating multidimensional voiceprint data, Z-score standardization or Min-Max normalization can be performed on each feature dimension to eliminate dimensional differences and improve the convergence speed and stability of subsequent classification models.

[0055] In specific implementation scenarios, based on the above solutions, a low-complexity feature extraction process can be designed to meet the deployment needs of edge devices, such as reducing the MFCC dimension and simplifying the PNCC calculation path, to ensure real-time operation even in resource-constrained environments.

[0056] Based on the prototype network, the voiceprint feature vectors of known fault types are abstracted into prototype vectors to form a support set. The extracted multidimensional voiceprint features are converted into feature vectors. By calculating the Euclidean distance between the feature vectors and each prototype vector, the similarity between the feature vectors and known fault categories is determined. The corresponding fault category is matched according to the minimum distance value, and the preliminary classification result is output.

[0057] As mentioned above, firstly, using previously accumulated labeled data, multi-dimensional acoustic signature data for various typical faults (such as gear wear, bearing pitting, coupling looseness, generator winding abnormalities, etc.) and normal operating conditions are collected. For each type of fault, several verified acoustic signature samples are selected, and their corresponding multi-dimensional acoustic signature feature vectors are aggregated, typically using the mean or weighted average method, to generate the "prototype vector" for that category. The prototype vectors of all categories together constitute the "support set," serving as the knowledge foundation for the classification model.

[0058] Once newly acquired voiceprint data enters the system, it undergoes the aforementioned preprocessing and feature extraction steps to generate a feature vector for the current moment. The system then performs a similarity measurement between this vector and each prototype vector in the support set, using Euclidean distance as the metric. The smaller the Euclidean distance, the closer the current feature vector is to the typical pattern of that fault category, and the higher the similarity.

[0059] The system then compares the distance between the current feature vector and all prototype vectors, selects the class with the smallest distance as the matching result, determines the most likely fault category corresponding to the audio segment, and outputs a "preliminary classification result". This process does not require complex neural network inference, has high computational efficiency, is suitable for deployment on edge devices, and can maintain a high recognition accuracy even when there are only a few samples (e.g., 5–20) per class.

[0060] The advantages of this mechanism are: it does not rely on a large number of training samples, thus avoiding the risk of overfitting in traditional deep learning models; the model structure is simple and updates are flexible; it supports incremental learning, which facilitates the subsequent expansion of new fault types.

[0061] For example, taking a certain type of turbine in a wind farm as an example, the system has accumulated the following four types of labeled samples: Normal operating status (15 samples) Slight wear on the gearbox (8 samples) Pitting on the outer ring of the main bearing (6 samples) Generator cooling fan loose (7 samples) After standardizing the multidimensional voiceprint feature vector (e.g., 29-dimensional) of each sample class, the feature mean of the same sample class is calculated to obtain four prototype vectors, which constitute the support set.

[0062] Given a newly acquired audio signal, after preprocessing and feature extraction, a 29-dimensional feature vector is generated. The system sequentially calculates the Euclidean distances between this vector and four prototype vectors, with the following results: Distance from the "normal" prototype: 0.42 Distance from the "gear wear" prototype: 0.38 Distance from the prototype of "bearing pitting": 0.61 Distance from the "loose fan" prototype: 0.73 The minimum distance was 0.38, corresponding to the "minor gearbox wear" category. Therefore, the system determined that this audio clip most likely reflected early wear in the gearbox and output the "preliminary classification result" as "minor gearbox wear".

[0063] The results are consistent with the trend of vibration data from the on-site CMS system, validating the classification effectiveness. Notably, although the "gear wear" class was constructed using only 8 samples, it could still accurately identify new samples, demonstrating the effectiveness of small-sample learning.

[0064] It should be noted that, in specific implementation scenarios, based on the above scheme, different weights can be assigned to samples in the support set according to the sample confidence level or collection time. For example, recent samples have higher weights, or samples that have been manually verified have higher weights, thereby improving the representativeness and timeliness of the prototype vector.

[0065] In specific implementation scenarios, in addition to Euclidean distance, Mahalanobis distance (considering feature covariance structure) or cosine similarity (focusing on directional consistency) can be introduced on the basis of the above scheme to select the optimal measurement method under specific fault types and improve classification accuracy.

[0066] In specific implementation scenarios, based on the above solutions, a two-level classification mechanism can be constructed for scenarios with many types of faults: the first level determines the major categories (such as mechanical and electrical), and the second level performs fine-grained matching in the subcategories to reduce the false positive rate.

[0067] In specific implementation scenarios, in addition to the above solutions, a confidence score (such as the ratio of the minimum distance to the second minimum distance) can be added to the output of the preliminary classification results, so that the operation and maintenance personnel can determine whether manual review is required, thereby improving the availability of the system.

[0068] In specific implementation scenarios, based on the above solutions, when the scale of the supported set increases, cluster analysis can be performed on samples of the same type, retaining only the cluster centers as prototypes, or an incremental prototype update strategy can be adopted to avoid excessive consumption of storage and computing resources.

[0069] In specific implementation scenarios, based on the above scheme, it is also possible to introduce the prototype vectors of other wind turbines of the same model as prior knowledge, and fine-tune the prototype with a small number of local samples to achieve a rapid deployment mode of "one machine training and multiple machines sharing".

[0070] In specific implementation scenarios, based on the above scheme, a sliding window statistical analysis can be performed on the classification results of multiple consecutive frames. If most frames point to the same category, the state is confirmed, reducing misjudgments caused by transient interference.

[0071] In specific implementation scenarios, based on the above scheme, the preliminary classification results can not only be used for normal / fault judgment, but their minimum distance value can also be used as the input for the subsequent "anomaly identification" step, realizing data closure between modules.

[0072] Based on the preliminary classification results, a preset threshold is used to determine whether the feature vector belongs to the abnormal category; if it is marked as abnormal, the abnormal multidimensional voiceprint data is stored in the dynamically updated abnormal voiceprint knowledge base.

[0073] As mentioned above, based on the initial classification, a distance threshold mechanism is introduced to determine whether the current voiceprint features significantly deviate from all known fault modes, thereby identifying potential abnormal states.

[0074] Specifically, while the prototype network outputs preliminary classification results, the system records the minimum similarity distance (e.g., Euclidean distance) between the current feature vector and the nearest prototype vector. This distance reflects the degree of matching between the current voiceprint and the closest known fault category. If the distance is less than a preset threshold, it indicates that the current state is highly similar to a certain known category and can be considered to belong to the normal or known fault category; conversely, if the distance exceeds the preset threshold, it indicates that the current voiceprint features differ significantly from all known categories, and may very well represent an abnormal state not covered by the support set, such as a novel fault, a compound fault, a sensor malfunction, or a sudden change in equipment structure.

[0075] At this point, the system marks the multidimensional voiceprint data corresponding to the feature vector as "abnormal" and stores it in a dynamically updated abnormal voiceprint knowledge base. This knowledge base is independent of the support set and is specifically used to store suspicious acoustic patterns discovered during system operation, forming a long-term memory and accumulation mechanism for unknown anomalies.

[0076] This abnormal voiceprint knowledge base has the following characteristics: Structured storage: Each piece of abnormal data is accompanied by metadata such as timestamp, collection location, operating conditions (e.g., wind speed, power), preliminary classification results, and distance value, which facilitates subsequent analysis; Deduplication and aggregation: Merging similar recurring abnormal patterns to avoid data redundancy; Priority labeling: Ranking abnormal samples based on indicators such as frequency of occurrence and distance value, and prioritizing the processing of high-frequency or high-deviation samples; Human interaction interface: Supports operation and maintenance personnel to replay, annotate and confirm abnormal data, providing a basis for subsequent model updates.

[0077] Through this mechanism, the system can not only identify known faults, but also proactively discover potential anomalies, realizing the transformation from "passive response" to "proactive early warning", and significantly improving the intelligence level and adaptability of the diagnostic system.

[0078] For example, let's take the continuous operation monitoring of a certain wind turbine as an example: The system supports three types of prototype vectors: "normal", "gear wear" and "bearing pitting". The preset distance threshold is 0.5 (determined based on historical normal sample statistics).

[0079] An audio clip collected on a certain day was processed, and its feature vector was 0.62 away from the nearest prototype ("normal"), which exceeded the threshold of 0.5.

[0080] Although the initial classification result output by the system was "normal", the voiceprint was still judged as "abnormal" because the distance exceeded the standard, and its multi-dimensional voiceprint data, along with the timestamp (April 10, 2025, 14:30), wind speed (9.2m / s), power (1.8MW), and other information, were stored in the abnormal voiceprint knowledge base.

[0081] Over the next three days, the same anomaly occurred five times consecutively, with highly similar voiceprint patterns. Maintenance personnel retrieved data from the knowledge base for manual analysis and confirmed it was caused by "slight resonance of the cooling fan," a novel fault mode.

[0082] The anomalous sample was labeled as "fan resonance" and added to the support set, triggering a model update. From then on, the system could identify this type of fault.

[0083] This example demonstrates that even if the initial classification result is "normal", the system can still detect potential anomalies through the distance threshold mechanism, avoiding missed reports, which reflects the key role of this mechanism in early warning.

[0084] It should be noted that, in specific implementation scenarios, the threshold can be preset to dynamically change with time or operating conditions, based on the above scheme. For example, the threshold can be appropriately relaxed in high wind speed ranges to avoid false alarms, while the threshold can be tightened during low load operation to improve sensitivity and enhance system adaptability.

[0085] In specific implementation scenarios, in addition to the minimum distance, composite judgment conditions can be constructed by combining indicators such as the second smallest distance, distance ratio, and confidence score to reduce the risk of misjudgment caused by a single threshold.

[0086] In specific implementation scenarios, based on the above solutions, the data in the abnormal voiceprint knowledge base can be automatically clustered to identify multiple potential new fault modes, assisting manual classification and labeling.

[0087] In specific implementation scenarios, based on the above solutions, anomaly markers can be correlated with parameters such as temperature, vibration, and current in the SCADA system to verify the physical meaning of the anomalies and improve diagnostic reliability.

[0088] In specific implementation scenarios, based on the above scheme, the frequency of anomalies in the same unit can be statistically analyzed. If the number of anomalies increases significantly within a unit time, a trend warning can be triggered to indicate the development of potential faults.

[0089] In specific implementation scenarios, based on the above solutions, multi-level user permission settings can be supported to ensure that only authorized personnel can modify or delete abnormal data; at the same time, the history of knowledge base changes is recorded for easy traceability and auditing.

[0090] In specific implementation scenarios, based on the above solutions, it is also possible to allow the abnormal soundprint knowledge base of multiple wind turbine units to be centrally managed in the cloud, realize the sharing of abnormal modes across units, and accelerate the identification and response to new types of faults.

[0091] In specific implementation scenarios, based on the above solution, when the cumulative number of a certain type of abnormal samples reaches a set number (such as 5 times) and the pattern consistency is high, the system can automatically notify the operation and maintenance personnel to mark it, forming a complete closed loop of "discovery-accumulation-confirmation-learning".

[0092] In specific implementation scenarios, based on the above solutions, to adapt to edge device deployment, abnormal voiceprint data can be compressed and stored, retaining only key feature vectors and metadata to reduce storage overhead.

[0093] By manually labeling frequently occurring anomalous samples and adding them to the support set, the prototype network can be retrained or adjusted to update the prototype vector.

[0094] As mentioned above, suspicious data identified and stored in the abnormal voiceprint knowledge base in the early stage are transformed into effective training samples after being manually confirmed, thereby updating the knowledge system of the classification model, enabling the system to "run, learn, and optimize at the same time".

[0095] Specifically, during the long-term operation of the system, the abnormal voiceprint knowledge base will continuously accumulate multi-dimensional voiceprint data marked as "abnormal". This data represents acoustic behavior that deviates significantly from known fault modes, and may correspond to new faults, compound faults, or performance degradation caused by equipment aging. However, automatically determined "abnormalities" may contain false alarms (such as transient interference or sudden environmental changes), so a manual annotation process is required for verification.

[0096] Maintenance personnel or domain experts analyze and confirm frequently occurring and stable anomaly samples by playing back audio, combining SCADA data, and on-site inspection information. Once it is confirmed that the sample represents a new categorizable fault mode (such as "yaw motor slippage" or "hydraulic system leakage"), the sample is categorized.

[0097] Subsequently, these manually labeled outlier samples are added as new positive samples to the original support set, forming an expanded support set. The support set is the knowledge base of the prototype network classification model, and its updates directly determine the boundary of the model's recognition ability.

[0098] After the support set is updated, the system retrains or adjusts the prototype network. "Retraining" refers to recalculating the prototype vectors of each category (including newly added categories) based on the expanded support set, for example, by generating a new prototype by taking the mean of the feature vectors of samples of the same type. "Adjustment" refers to incrementally updating only the affected prototype vectors without completely rebuilding the model, in order to maintain model stability and reduce computational overhead.

[0099] Ultimately, the updated prototype vector was deployed to the online identification module, enabling the system to recognize this new type of fault. This process achieves a complete closed loop from "anomaly detection" to "fault confirmation" and then to "model evolution," significantly improving the system's intelligence level and long-term applicability.

[0100] For example, let's take a certain type of turbine in a wind farm as an example: In the past month, the system has recorded a total of 8 “abnormal” events with similar acoustic characteristics in the abnormal voiceprint knowledge base. The voiceprints show periodic fluctuations in energy in the mid-to-high frequency range, and the preliminary classification is far from normal and known faults.

[0101] Maintenance personnel retrieved these abnormal data and, combined with the SCADA data from the same period, observed fluctuations in hydraulic oil pressure and delays in actuator response. After on-site inspection, they confirmed that the problem was caused by "micro-leakage in the hydraulic braking system," a novel fault that had not been previously recorded.

[0102] The maintenance personnel uniformly labeled these 8 abnormal samples as "hydraulic system micro-leakage" and added them to the support set, forming a new support set that includes the original three types of faults and the newly added categories.

[0103] The system recalculates the prototype vectors of the four types of faults based on the new support set, where the prototype of the "hydraulic system micro-leakage" type is obtained from the feature values ​​of 8 labeled samples.

[0104] The updated prototype network is deployed to edge devices, and when similar faults occur in the future, the system can directly identify and issue an alarm without having to rely on the anomaly detection mechanism again.

[0105] This example demonstrates that through manual annotation and model updates, the system successfully transformed "unknown anomalies" into "known faults," achieving self-evolution of diagnostic capabilities.

[0106] It should be noted that, in specific implementation scenarios, the system can automatically sort abnormal samples based on the above solution, prioritizing data that occurs frequently, has large distance values, and represents typical working conditions for manual annotation, thereby improving annotation efficiency.

[0107] In specific implementation scenarios, based on the above scheme, for the supplementation of existing category samples (such as adding normal samples), the prototype vector can be dynamically adjusted by using a sliding window or exponential weighted average method to avoid full retraining and improve update efficiency.

[0108] In specific implementation scenarios, based on the above scheme, when there are few new fault category samples (such as less than 5), transfer learning can be used to borrow prototypes of similar equipment or components as initial values, and then fine-tune them with a small number of local samples.

[0109] In specific implementation scenarios, based on the above approach, a new version can be generated with each prototype vector update, supporting the saving of historical versions and performance comparison. If the recognition accuracy of the new model decreases, it can be quickly rolled back to a stable version, ensuring system reliability.

[0110] In specific implementation scenarios, based on the above scheme, multiple units of the same model can be allowed to share the annotation results and the updated prototype vector, realizing a swarm intelligence learning mode of "one unit discovers, multiple units benefit".

[0111] In specific implementation scenarios, update trigger rules can be set based on the above scheme, such as "the same abnormal pattern is labeled more than 3 times" or "the cumulative number of newly added category samples reaches 10". Once the conditions are met, the model update process will be started automatically, reducing manual intervention.

[0112] In specific implementation scenarios, lightweight prototype updates can also be implemented on edge devices based on the above solutions. Only local computing resources are needed to complete the mean calculation and vector replacement, without the need for deep training in the cloud, thus meeting real-time requirements.

[0113] In specific implementation scenarios, based on the above solutions, after updates, the system can evaluate the performance of the new model through historical test sets or cross-validation to ensure that the classification accuracy is not lower than a preset threshold (such as 90%) before it can be officially launched.

[0114] In specific implementation scenarios, based on the above solutions, for fault categories that have not appeared for a long time, an aging strategy can be set to reduce their prototype vector weights or temporarily freeze them to prevent the model from overexpanding and keep the knowledge system concise and efficient.

[0115] According to one embodiment of this application, in the step of decomposing the original audio signal into multiple intrinsic mode components by variational mode decomposition, the number of intrinsic mode components is set to 3-7, and the target mode is selected according to the following conditions: The intrinsic mode components that match the vibration frequency of key components of the wind turbine were selected, with the center frequency range being 50Hz-2000Hz. Retain intrinsic mode components with energy density higher than a preset threshold, and remove low-energy noise components; Intrinsic mode components are screened using kurtosis indices, retaining those with kurtosis values ​​greater than 1.5 to preserve non-Gaussian features in the signal.

[0116] As mentioned above, in the step of decomposing the original audio signal into multiple intrinsic mode components (IMFs) using variational mode decomposition (VMD), the first step is to determine the number of modes to be decomposed. The number of IMFs is typically set to 3 to 7, a range determined based on the complexity of the wind turbine audio signal and actual operating conditions. Too few modes will result in insufficient separation of different frequency components, leading to mode aliasing; too many modes may over-decompose signals from the same physical source, introducing redundant components and increasing the burden on subsequent processing. A setting of 3 to 7 modes ensures sufficient signal decomposition while balancing computational efficiency and the interpretability of the physical meaning.

[0117] After obtaining multiple intrinsic mode components, it is necessary to screen out the target modes that are relevant to the operating state of the wind turbine, and eliminate invalid components mainly composed of environmental noise or interference. The screening process is based on the following three conditions: Intrinsic modal components that match the vibration frequencies of key wind turbine components are selected. Key components of wind turbines, such as main bearings, gearboxes, and generators, generate vibration signals within specific frequency ranges during operation, which propagate through the structure to form collectable acoustic characteristics. The typical vibration frequencies of these components are mostly concentrated in the range of 50Hz to 2000Hz. Therefore, spectral analysis is performed on each intrinsic modal component to calculate its center frequency. Only modes with center frequencies falling within the 50Hz–2000Hz range are retained to ensure that the selected modes reflect the actual operating state of the core components of the equipment.

[0118] Intrinsic mode components with energy densities exceeding a preset threshold are retained. Energy density reflects the contribution of a particular mode to the overall signal energy. Under normal operating conditions, acoustic signals related to mechanical vibrations typically have high energy concentration, while background signals such as environmental noise and electromagnetic interference have lower energy and are more diffusely distributed. Therefore, a preset energy density threshold is set, for example, 1.2 times the average energy of all modes or 15% of the total signal energy. Only modes with energy densities exceeding this threshold are retained, effectively eliminating weak noise components that do not contribute to fault identification.

[0119] The intrinsic mode components are further screened using the kurtosis index. Kurtosis is a statistical measure of the non-Gaussianity of the probability distribution of a signal. High kurtosis values ​​usually correspond to the presence of impulsive or sudden components in the signal. These characteristics are often closely related to transient vibrations caused by early mechanical faults (such as pitting, cracks, loosening, etc.). Therefore, the kurtosis value of each mode is calculated, and only components with a kurtosis greater than 1.5 are retained. This threshold can effectively distinguish between stationary background noise (kurtosis close to 1.0) and non-stationary signals containing fault characteristics, thereby enhancing the sensitivity of the target mode to early faults.

[0120] Based on the above three screening criteria, the intrinsic modal components were evaluated sequentially from three dimensions: frequency characteristics, energy contribution, and statistical characteristics, ultimately identifying several target modes. These target modes were then linearly superimposed to reconstruct a new audio signal. This reconstructed signal retains the main acoustic components related to the operating status of key wind turbine components, while significantly suppressing irrelevant noise and interference, providing a high-quality input signal for subsequent feature extraction and fault identification.

[0121] According to one embodiment of this application, in the step of calculating the Mel frequency cepstral coefficients of the denoised audio signal, the calculation parameters of the Mel frequency cepstral coefficients include: The audio signal is divided into time-domain segments with a frame length of 20ms and a frame shift of 10ms; It employs 24-40 Mel filters, covering a frequency range of 20Hz-4000Hz; After taking the logarithm of the power spectrum, a discrete cosine transform is performed to extract the 12-20 dimension Mel frequency cepstral coefficients, and the first 12 dimensions are retained as the main features.

[0122] As described above, in the step of calculating the Mel frequency cepstral coefficients (MFCC) of the denoised audio signal, the denoised audio signal obtained through preprocessing methods such as variational mode decomposition is first divided into short time frames for time-frequency analysis. Since audio signals have non-stationary characteristics, they need to be segmented into a series of short time periods to approximate the stationarity assumption. Specifically, the audio signal is segmented into 20-millisecond frame lengths, with adjacent frames sliding by a 10-millisecond frame shift to achieve frame overlap. This setting ensures time resolution while avoiding a decrease in frequency resolution due to excessively short frame lengths, balancing the dynamics of signal changes with the accuracy of spectral estimation.

[0123] After framing, each frame of signal is windowed (usually using a Hamming window) to reduce spectral leakage caused by truncation. Then, a Fast Fourier Transform (FFT) is performed to convert the time-domain signal to the frequency-domain signal, obtaining the power spectrum of that frame. Next, a Mel filter bank is applied in the frequency domain. The Mel filter bank simulates the nonlinear characteristics of human hearing's perception of different frequencies, mapping a linear frequency scale to the Mel scale. This scheme uses 24 to 40 triangular bandpass filters, evenly distributed across the Mel scale, covering a frequency range of 20 Hz to 4000 Hz. This frequency range encompasses the typical acoustic signal bands generated by the main mechanical components of the wind turbine (such as gears and bearings), while avoiding extremely low-frequency environmental disturbances and high-frequency wind noise interference.

[0124] Each filter performs a weighted integral of the power spectrum to obtain the energy value for the corresponding frequency band, forming the output of the filter bank. Then, the logarithm of the outputs of all filters is taken to compress the dynamic range and highlight low-to-medium energy components, enhancing sensitivity to weak signals. The logarithmic energy value reflects the energy distribution characteristics of each Mel frequency band and is a key intermediate result in MFCC calculation.

[0125] Finally, a Discrete Cosine Transform (DCT) is performed on the logarithmic filter bank energy to convert the frequency domain correlation into a time-domain coefficient representation. DCT yields 12 to 20-dimensional Mel-frequency cepstral coefficients. The first few dimensions primarily reflect the envelope information of the spectrum, closely related to the sound source characteristics, while higher-dimensional coefficients correspond to detail fluctuations and are susceptible to noise. Therefore, in practical applications, retaining the first 12 dimensions of MFCC coefficients as the main features effectively characterizes the spectral profile of the audio signal while reducing the feature dimensionality, thus improving the computational efficiency and robustness of subsequent classification models.

[0126] In summary, by setting a frame length of 20ms, a frame shift of 10ms, and 24–40 Mel filters covering the frequency range of 20Hz–4000Hz, and extracting the first 12 dimensions of MFCC coefficients, a standardized feature extraction process suitable for wind turbine voiceprint recognition was constructed, ensuring the stability, comparability, and physical interpretability of the features.

[0127] According to one embodiment of this application, in the step of enhancing feature robustness by combining the perceptual noise robustness coefficient algorithm, the implementation of the perceptual noise robustness coefficient includes: The power spectral density of noise is estimated from environmental noise samples and separated from the spectrum of the audio signal; The perceived noise robustness coefficients are weighted, and the weighting coefficients are correlated with the signal-to-noise ratio of the noise. The compensation parameters are dynamically adjusted based on real-time noise changes to ensure that the perceived noise robustness coefficient remains stable in complex noise environments.

[0128] As described above, in the step of enhancing feature robustness by incorporating the Perceived Noise Robustness Coefficient (PNCC) algorithm, the background noise characteristics are first modeled using environmental noise samples to improve the stability of the features under complex operating conditions. Specifically, environmental noise signals are collected for a period of time as noise samples when the wind turbine is in a state without significant mechanical noise sources, such as when it is shut down or unloaded. Based on these samples, the average power spectral density is calculated to obtain the energy distribution model of the noise at each frequency component. This noise model is used for subsequent compensation processing of the audio signal spectrum under operating conditions.

[0129] When extracting features from the denoised audio signal, its spectrum is compared with the pre-estimated noise power spectral density. Using methods such as spectral subtraction or minimum mean square error estimation, noise-affected components are separated from the original signal spectrum and suppressed or compensated for. This process effectively reduces the influence of non-target sound sources such as wind noise, rain noise, and electromagnetic interference, highlighting the acoustic characteristics of the equipment itself.

[0130] Furthermore, the perceived noise robustness coefficients are weighted. Since different frequency bands are affected by noise to varying degrees, and the signal-to-noise ratio (SNR) changes with frequency and operating conditions, weighting coefficients related to SNR are introduced to adjust the characteristic components of each frequency band. Specifically, in frequency bands with low SNR (such as high-frequency wind noise regions), smaller weights are assigned to reduce their contribution to the overall characteristics; while in frequency bands with high SNR and concentrated effective signals (such as mid-to-low frequency mechanical vibration regions), larger weights are assigned to enhance their characterization ability. This weighting mechanism allows the PNCC characteristics to adaptively focus on frequency bands with high reliability, improving the overall robustness of the characteristics.

[0131] Furthermore, considering the time-varying nature of noise environments at wind power sites (e.g., fluctuations in wind noise intensity due to changes in wind speed), the system dynamically monitors changes in noise levels based on real-time acquired audio signals and adjusts the compensation parameters in the PNCC algorithm accordingly. For example, when an increase in background noise energy is detected, the spectral compensation intensity is automatically increased or the noise model is updated; when the noise tends to stabilize, the current parameters are maintained to avoid over-correction. This dynamic adjustment mechanism ensures that PNCC features remain consistent under different environmental conditions, avoiding feature drift or classification performance degradation due to sudden noise changes.

[0132] In summary, noise separation is achieved through noise power spectral density estimation, combined with signal-to-noise ratio-related weighted processing, and compensation parameters are dynamically adjusted according to real-time noise changes. This enables the perceived noise robustness coefficient to effectively enhance the anti-interference capability of acoustic signature features in complex and non-stationary wind power operating environments, providing a more stable and reliable input for subsequent fault identification.

[0133] According to one embodiment of this application, in the step of extracting the statistical features of spectral centroid, spectral entropy, energy, kurtosis, and zero-crossing rate, the calculation method for each feature is as follows: The formula is calculated by weighted averaging of spectral energy. , where f i For the frequency point, E i For the corresponding energy; The uncertainty of the spectral distribution is calculated based on the Shannon entropy formula, which characterizes the complexity of the signal; The number of times an audio signal passes through zero within a unit of time is counted to reflect the distribution characteristics of high-frequency components.

[0134] As mentioned above, in the steps of extracting statistical features such as spectral centroid, spectral entropy, energy, kurtosis and zero-crossing rate, each feature describes the dynamic behavior of the wind turbine's operating audio signal from the perspectives of frequency domain, time domain and signal distribution characteristics, forming a multi-dimensional representation of the equipment status.

[0135] The spectral centroid is calculated using a weighted average of spectral energy. Specifically, after performing a Fourier transform on the audio signal to obtain its spectrum, the energy at each frequency point is used as a weight, multiplied by the corresponding frequency value, summed, and then divided by the total energy to obtain the spectral centroid. This value reflects the concentrated position of signal energy on the frequency axis. When key components of a wind turbine experience wear or failure, high-frequency components often increase, causing the spectral centroid to shift towards higher frequencies. Therefore, this characteristic has a good indicative role in early fault identification.

[0136] Spectral entropy, calculated based on the Shannon entropy formula, measures the degree of uncertainty or disorder in the spectral distribution of a signal. First, the spectral energy is normalized to a probability distribution. Then, the logarithm of the energy proportion for each frequency band is taken and summed with weights to obtain the spectral entropy value. A higher spectral entropy indicates a more dispersed energy distribution and a more complex signal; conversely, a lower spectral entropy indicates concentrated energy and a more stable signal. When equipment operates abnormally, such as due to loosening, impact, or non-stationary vibration, the spectral distribution typically becomes more chaotic, and the spectral entropy value increases accordingly. Therefore, this characteristic can be used to identify non-steady-state acoustic behavior.

[0137] Energy characteristics refer to the total or average energy of an audio signal in the time or frequency domain, typically obtained by summing the squares of the signal amplitudes within a frame or integrating the spectral energy. This characteristic reflects the overall strength of the acoustic signal and is closely related to equipment load and vibration amplitude. For example, when gear meshing is faulty or bearing damage worsens, mechanical vibration intensifies, leading to an increase in acoustic energy. Therefore, energy changes can serve as a reference indicator of the degree of fault development.

[0138] Kurtosis is the fourth-order statistical moment of a signal amplitude distribution, used to measure the presence of significant impulse components in a waveform. A higher kurtosis value indicates more sudden, sharp pulses in the signal, typically associated with early mechanical faults such as pitting, cracks, and localized spalling. Under normal operating conditions, audio signals approximate a Gaussian distribution with a kurtosis value close to 3 (or close to 0 after normalization). When periodic impulses or transient events occur, the kurtosis increases significantly, making this characteristic highly sensitive to minor faults.

[0139] Zero-crossing rate refers to the number of times an audio signal crosses zero level per unit time, typically counted frame by frame in the time domain. A higher zero-crossing rate indicates faster signal fluctuation and richer high-frequency components. This characteristic can be used to distinguish between stable operation and abnormal states related to high-frequency noise such as friction and flutter. For example, when ice forms on the surface of a wind turbine blade or when aerodynamic disturbances are present, high-frequency airflow noise may be generated, leading to a significant increase in the zero-crossing rate.

[0140] The five statistical features mentioned above, together with features such as Mel frequency cepstral coefficients, constitute multidimensional voiceprint data, which supplements the physical characteristics of audio signals from different perspectives, enhances the ability of feature vectors to represent the operating status of equipment, and can maintain good distinguishability and stability, especially in complex noise environments.

[0141] According to one embodiment of this application, in the step of abstracting the voiceprint feature vector of a known fault type into a prototype vector based on a prototype network, the method for generating the prototype vector includes: Z-score standardization is applied to the feature vectors supporting the same type of fault in the set, so that their mean is 0 and their variance is 1. The prototype vector is generated by taking the mean of the feature vectors of the same type. If the number of samples of the same type is insufficient, the K-nearest neighbor weighted average method is used to supplement it. If the eigenvector dimension is higher than 50, principal component analysis is used to reduce the dimension to 20-40 to reduce computational complexity.

[0142] As described above, in the step of abstracting the acoustic feature vectors of known fault types into prototype vectors based on the prototype network, multiple acoustic feature vectors belonging to the same fault category are first obtained from the support set. These feature vectors originate from previously collected and labeled multidimensional acoustic data, and each vector represents the acoustic pattern of the fault type under specific operating conditions. To improve the accuracy and stability of subsequent similarity calculations, Z-score standardization is first applied to all feature vectors of the same category in the support set. This process subtracts the mean of that dimension from the samples of the same category, and then divides by the standard deviation, so that the processed data has a mean of 0 and a variance of 1 in each dimension. This operation eliminates the differences in dimensions and numerical ranges between different feature dimensions, preventing high-amplitude features from dominating distance calculations, thereby ensuring a more balanced contribution from various features.

[0143] After standardization, a prototype vector for each fault category is generated by calculating the arithmetic mean of feature vectors of the same category. The prototype vector represents the "central pattern" of the fault category and serves as the benchmark for similarity comparison during classification. When the number of samples for a category is sufficient (e.g., at least 5), a stable prototype can be obtained by directly taking its mean. However, in actual wind power operation and maintenance, some fault types may only have a small number of usable samples, resulting in insufficient representativeness of the prototype vector. Therefore, if the number of samples of the same category is insufficient, the K-nearest neighbor weighted average method is used to supplement it. Specifically, K closest samples (based on distance in the feature space) are selected from other similar fault categories, and weighted according to their similarity to the current category, and then used to participate in the calculation of the prototype vector. This method effectively expands the effective data under small sample conditions, improving the robustness and generalization ability of the prototype vector.

[0144] Furthermore, when the original feature vector has a high dimensionality (e.g., exceeding 50 dimensions), directly calculating the distance incurs significant computational overhead and may introduce redundant or noisy dimensions, impacting classification performance. Therefore, if the feature vector dimension exceeds 50 dimensions, Principal Component Analysis (PCA) is introduced for dimensionality reduction. PCA projects the original high-dimensional features into a new low-dimensional space through a linear transformation, retaining the top 20 to 40 principal components with the largest variance contributions, thus significantly reducing data dimensionality while preserving as much original information as possible. The dimensionality-reduced feature vectors are used to generate prototype vectors and subsequent similarity matching, effectively reducing computational complexity and improving the model's efficiency on edge devices.

[0145] In summary, by employing Z-score standardization, mean aggregation and K-nearest neighbor supplementation, and principal component analysis for dimensionality reduction, the accuracy, stability, and computational feasibility of the prototype vectors under conditions of small samples, high dimensionality, and multiple sources are ensured, providing a reliable support set foundation for distance-based fault identification.

[0146] According to one embodiment of this application, the preset threshold is set based on the Euclidean distance distribution under normal conditions, and is dynamically adjusted through the following steps: The 95th percentile was calculated based on the Euclidean distance from historical normal samples and used as the initial threshold. Based on the Euclidean distance distribution of real-time operational data, the threshold is updated every 24 hours, and the correction formula is as follows: ; Where μ is the current mean, σ is the standard deviation, and k is the empirical coefficient.

[0147] As described above, in setting and dynamically adjusting the preset threshold, the initial threshold is first determined based on the Euclidean distance distribution of historical normal samples. Specifically, a certain number of samples are randomly selected from a large amount of voiceprint data labeled as normal as a historical normal sample set. After feature extraction and standardization of these samples, the Euclidean distance between each pair of samples is calculated, forming a distance distribution that reflects the degree of similarity of voiceprint features under normal conditions. Then, statistical analysis is performed on this distance distribution, and its 95th percentile is calculated as the initial threshold. This threshold indicates that under normal conditions, the distance between 95% of sample pairs is less than or equal to this value, and only 5% of sample pairs have a distance exceeding this range. Therefore, it can be considered that when the distance between the real-time monitored voiceprint features and the normal pattern exceeds this threshold, there is a high probability of an anomaly.

[0148] During model operation, the threshold is dynamically adjusted based on real-time collected operational data to adapt to changes in equipment operating conditions and fluctuations in environmental noise. Specifically, the system collects a new batch of operational data every 24 hours, performs feature extraction and standardization on it, and calculates the Euclidean distance between each sample in this batch of data and the current normal mode prototype vector, forming a new distance distribution. Based on this distribution, its mean μ and standard deviation σ are calculated, representing the average similarity and dispersion of the voiceprint features under the current operational state, respectively.

[0149] To dynamically adjust the threshold, a correction formula is introduced: New Threshold = μ + k * σ, where k is an empirical coefficient, typically between 1.5 and 3, used to balance sensitivity and robustness. When k is small, the new threshold is closer to the current mean, making it more sensitive to anomaly detection, but may increase the false alarm rate; when k is large, the new threshold deviates from the mean, increasing the confidence of the judgment, but may lead to missed detections. Through experimental verification and parameter optimization, a suitable k value is selected so that the threshold can both promptly capture fault signals that significantly deviate from the normal pattern and avoid frequent alarms caused by small fluctuations.

[0150] In summary, an adaptive anomaly detection mechanism was constructed by determining the initial threshold based on the 95th percentile of historical normal samples and dynamically adjusting the threshold by combining the mean and standard deviation of real-time operating data. This mechanism can continuously track the changing trend of acoustic signature characteristics during the long-term operation of wind turbine units, automatically update the judgment benchmark, and ensure the accuracy and timeliness of fault identification.

[0151] According to one embodiment of this application, in the step of determining similarity by calculating the Euclidean distance between the feature vector and each prototype vector, the calculation of the Euclidean distance includes: Normalize the Euclidean distance to the [0,1] interval based on the maximum and minimum values; Weights are assigned based on the importance of the feature dimensions; If the minimum distance is less than the preset threshold, it is considered a match; otherwise, it is marked as an anomaly.

[0152] As described above, in the step of determining similarity by calculating the Euclidean distance between the feature vector and each prototype vector, the Euclidean distance between the voiceprint feature vector to be identified and the prototype vector of each known fault category in the support set is first calculated. The Euclidean distance reflects the straight-line distance between two vectors in multidimensional space; the smaller the distance, the more similar the two are, and the higher the degree of matching.

[0153] To eliminate the influence of differences in the dimensions of different feature dimensions and the overall amplitude range on distance calculation, the calculated original Euclidean distance is normalized. Specifically, the maximum-minimum normalization method is used, which involves subtracting the minimum value among all distances from the current distance, and then dividing by the difference between the maximum and minimum values, mapping the result to the [0,1] interval. The normalized distance value has a uniform scale, which facilitates the subsequent setting of a uniform judgment threshold and cross-class comparisons.

[0154] To further improve classification accuracy, a weighting mechanism is introduced, taking into account the actual contribution of each feature dimension to fault identification. Since different features (such as MFCC, spectral entropy, kurtosis, etc.) have varying importance in distinguishing specific fault types, a corresponding weight coefficient is assigned to each feature dimension based on the feature importance analysis results (which can be determined through training with historical data or expert experience). Before calculating the Euclidean distance, each dimension of the feature vector and prototype vector is weighted, ensuring that key features have a greater weight in the distance calculation, thereby enhancing the classification model's responsiveness to sensitive features.

[0155] After calculating the weighted and normalized Euclidean distance, the system selects the smallest distance value as the optimal matching result between the current feature vector and the known categories. If the minimum distance is less than a preset threshold, the feature vector is determined to belong to a certain known fault category, and the corresponding preliminary classification result is output. If the minimum distance is greater than or equal to the preset threshold, it indicates that the current voiceprint feature differs significantly from all known patterns and cannot be reliably matched. The system marks it as "abnormal" and triggers the subsequent abnormal handling process.

[0156] In summary, by ensuring the consistency of distance measurement through normalization, enhancing the discrimination capability of key dimensions through feature weighting, and combining it with preset thresholds for matching and judgment, the system achieves accurate identification and anomaly detection of wind turbine operating status under complex conditions, thereby enhancing the robustness and practicality of the entire acoustic fingerprint fault diagnosis system.

[0157] A second aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any of the embodiments of the first aspect above.

[0158] Figure 2 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 2 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 may call logical instructions in the memory 830 to execute the method in any of the embodiments of the first aspect described above, the method including: The original audio signal of the wind turbine is collected during operation, and the original audio signal is decomposed into multiple intrinsic mode components through variational mode decomposition. The intrinsic mode components are screened according to the frequency distribution, energy density and kurtosis of the modes to obtain the target mode, so as to reconstruct the noise-reduced audio signal. Mel frequency cepstral coefficients are calculated for the denoised audio signal, and the feature robustness is enhanced by combining the perceptual noise robustness coefficient algorithm. At the same time, statistical features such as spectral centroid, spectral entropy, energy, kurtosis and zero-crossing rate are extracted to generate multidimensional voiceprint data. Based on the prototype network, the voiceprint feature vectors of known fault types are abstracted into prototype vectors to form a support set. The extracted multidimensional voiceprint features are converted into feature vectors. By calculating the Euclidean distance between the feature vectors and each prototype vector, the similarity between the feature vectors and known fault categories is determined. The corresponding fault category is matched according to the minimum distance value, and the preliminary classification result is output. Based on the preliminary classification results, a preset threshold is used to determine whether the feature vector belongs to the abnormal category; if it is marked as abnormal, the abnormal multidimensional voiceprint data is stored in the dynamically updated abnormal voiceprint knowledge base. By manually labeling frequently occurring anomalous samples and adding them to the support set, the prototype network can be retrained or adjusted to update the prototype vector.

[0159] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0160] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer being able to perform the methods provided by the above methods, the method comprising: The original audio signal of the wind turbine is collected during operation, and the original audio signal is decomposed into multiple intrinsic mode components through variational mode decomposition. The intrinsic mode components are screened according to the frequency distribution, energy density and kurtosis of the modes to obtain the target mode, so as to reconstruct the noise-reduced audio signal. Mel frequency cepstral coefficients are calculated for the denoised audio signal, and the feature robustness is enhanced by combining the perceptual noise robustness coefficient algorithm. At the same time, statistical features such as spectral centroid, spectral entropy, energy, kurtosis and zero-crossing rate are extracted to generate multidimensional voiceprint data. Based on the prototype network, the voiceprint feature vectors of known fault types are abstracted into prototype vectors to form a support set. The extracted multidimensional voiceprint features are converted into feature vectors. By calculating the Euclidean distance between the feature vectors and each prototype vector, the similarity between the feature vectors and known fault categories is determined. The corresponding fault category is matched according to the minimum distance value, and the preliminary classification result is output. Based on the preliminary classification results, a preset threshold is used to determine whether the feature vector belongs to the abnormal category; if it is marked as abnormal, the abnormal multidimensional voiceprint data is stored in the dynamically updated abnormal voiceprint knowledge base. By manually labeling frequently occurring anomalous samples and adding them to the support set, the prototype network can be retrained or adjusted to update the prototype vector.

[0161] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the cigarette box image recognition method provided by the methods described above, the method comprising: The original audio signal of the wind turbine is collected during operation, and the original audio signal is decomposed into multiple intrinsic mode components through variational mode decomposition. The intrinsic mode components are screened according to the frequency distribution, energy density and kurtosis of the modes to obtain the target mode, so as to reconstruct the noise-reduced audio signal. Mel frequency cepstral coefficients are calculated for the denoised audio signal, and the feature robustness is enhanced by combining the perceptual noise robustness coefficient algorithm. At the same time, statistical features such as spectral centroid, spectral entropy, energy, kurtosis and zero-crossing rate are extracted to generate multidimensional voiceprint data. Based on the prototype network, the voiceprint feature vectors of known fault types are abstracted into prototype vectors to form a support set. The extracted multidimensional voiceprint features are converted into feature vectors. By calculating the Euclidean distance between the feature vectors and each prototype vector, the similarity between the feature vectors and known fault categories is determined. The corresponding fault category is matched according to the minimum distance value, and the preliminary classification result is output. Based on the preliminary classification results, a preset threshold is used to determine whether the feature vector belongs to the abnormal category; if it is marked as abnormal, the abnormal multidimensional voiceprint data is stored in the dynamically updated abnormal voiceprint knowledge base. By manually labeling frequently occurring anomalous samples and adding them to the support set, the prototype network can be retrained or adjusted to update the prototype vector.

[0162] For any parts not mentioned in this application, existing technologies may be used or referenced.

[0163] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0164] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for identifying acoustic fingerprint faults in wind turbine generators, characterized in that, include: The original audio signal of the wind turbine is collected during operation, and the original audio signal is decomposed into multiple intrinsic mode components through variational mode decomposition. The intrinsic mode components are screened according to the frequency distribution, energy density and kurtosis of the modes to obtain the target mode, so as to reconstruct the noise-reduced audio signal. Mel frequency cepstral coefficients are calculated for the denoised audio signal, and the feature robustness is enhanced by combining the perceptual noise robustness coefficient algorithm. At the same time, statistical features such as spectral centroid, spectral entropy, energy, kurtosis and zero-crossing rate are extracted to generate multidimensional voiceprint data. Based on the prototype network, the voiceprint feature vectors of known fault types are abstracted into prototype vectors to form a support set. The extracted multidimensional voiceprint features are converted into feature vectors. By calculating the Euclidean distance between the feature vectors and each prototype vector, the similarity between the feature vectors and known fault categories is determined. The corresponding fault category is matched according to the minimum distance value, and the preliminary classification result is output. Based on the preliminary classification results, a preset threshold is used to determine whether the feature vector belongs to the abnormal category; If it is marked as abnormal, the abnormal multidimensional voiceprint data will be stored in the dynamically updated abnormal voiceprint knowledge base. By manually labeling frequently occurring anomalous samples and adding them to the support set, the prototype network can be retrained or adjusted to update the prototype vector.

2. The method according to claim 1, characterized in that, In the step of decomposing the original audio signal into multiple intrinsic mode components through variational mode decomposition, the number of intrinsic mode components is set to 3-7, and the target mode is selected according to the following conditions: The intrinsic mode components that match the vibration frequency of key components of the wind turbine were selected, with the center frequency range being 50Hz-2000Hz. Retain intrinsic mode components with energy density higher than a preset threshold, and remove low-energy noise components; Intrinsic mode components are screened using kurtosis indices, retaining those with kurtosis values ​​greater than 1.5 to preserve non-Gaussian features in the signal.

3. The method according to claim 1, characterized in that, In the step of calculating the Mel frequency cepstral coefficients of the denoised audio signal, the calculation parameters for the Mel frequency cepstral coefficients include: The audio signal is divided into time-domain segments with a frame length of 20ms and a frame shift of 10ms; It employs 24-40 Mel filters, covering a frequency range of 20Hz-4000Hz; After taking the logarithm of the power spectrum, a discrete cosine transform is performed to extract the 12-20 dimension Mel frequency cepstral coefficients, and the first 12 dimensions are retained as the main features.

4. The method according to claim 1, characterized in that, In the step of enhancing feature robustness by combining the perceptual noise robustness coefficient algorithm, the implementation of the perceptual noise robustness coefficient includes: The power spectral density of noise is estimated from environmental noise samples and separated from the spectrum of the audio signal; The perceived noise robustness coefficients are weighted, and the weighting coefficients are correlated with the signal-to-noise ratio of the noise. The compensation parameters are dynamically adjusted based on real-time noise changes to ensure that the perceived noise robustness coefficient remains stable in complex noise environments.

5. The method according to claim 1, characterized in that, In the steps of extracting the statistical features of spectral centroid, spectral entropy, energy, kurtosis, and zero-crossing rate, the calculation method for each feature is as follows: The formula is calculated by weighted averaging of spectral energy. , where f i For the frequency point, E i For corresponding energy; The uncertainty of the spectral distribution is calculated based on the Shannon entropy formula, which characterizes the complexity of the signal; The number of times an audio signal passes through zero within a unit of time is counted to reflect the distribution characteristics of high-frequency components.

6. The method according to claim 1, characterized in that, In the step of abstracting the voiceprint feature vectors of known fault types into prototype vectors based on prototype networks, the methods for generating prototype vectors include: Z-score standardization is applied to the feature vectors supporting the same type of fault in the set, so that their mean is 0 and their variance is 1. The prototype vector is generated by taking the mean of the feature vectors of the same type. If the number of samples of the same type is insufficient, the K-nearest neighbor weighted average method is used to supplement it. If the eigenvector dimension is higher than 50, principal component analysis is used to reduce the dimension to 20-40 to reduce computational complexity.

7. The method according to claim 1, characterized in that, The preset threshold is set based on the Euclidean distance distribution under normal conditions, and is dynamically adjusted through the following steps: The 95th percentile was calculated based on the Euclidean distance from historical normal samples and used as the initial threshold. Based on the Euclidean distance distribution of real-time operational data, the threshold is updated every 24 hours, and the correction formula is as follows: ； Where μ is the current mean, σ is the standard deviation, and k is the empirical coefficient.

8. The method according to claim 1, characterized in that, In the step of determining similarity by calculating the Euclidean distance between the feature vector and each prototype vector, the calculation of the Euclidean distance includes: Normalize the Euclidean distance to the [0,1] interval based on the maximum and minimum values; Weights are assigned based on the importance of the feature dimensions; If the minimum distance is less than the preset threshold, it is considered a match; otherwise, it is marked as an anomaly.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-8.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1-8.

Images