Unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism

Abstract

The present application relates to wind power equipment blade fault detection technology, and discloses an unsupervised wind power equipment blade fault detection method based on phase-aware parallel attention mechanism, which solves the problem that the existing wind power equipment blade fault detection method is strongly dependent on labeled data, has insufficient generalization ability under strong noise and variable working conditions, and is difficult to robustly capture weak transient fault signals and dynamic change characteristics.The present application scheme collects blade operation audio signals, extracts double-channel time-frequency characteristics containing amplitude spectrum and phase spectrum through improved short-time Fourier transform;A deep adversarial autoencoder is constructed using an encoder, a decoder and an auxiliary encoder containing a phase-aware parallel attention module, the model is optimized by reconstruction error loss, latent representation consistency loss, adversarial loss and phase consistency loss during offline training, and the normal working condition feature distribution is learned;In the inference stage, the fault is judged based on the feature distance score and the reconstruction error score.

Description

Technical Field

[0001] This invention relates to wind turbine blade fault detection technology, specifically to an unsupervised wind turbine blade fault detection method based on a phase-aware parallel attention mechanism. Background Technology

[0002] Blade fault detection in wind turbines is a critical aspect of wind power operation and maintenance, directly impacting the operational efficiency and safety of wind farms. Traditional blade fault detection methods primarily rely on periodic manual inspections and sensor monitoring. These methods are not only costly but also struggle to achieve real-time and comprehensive fault detection. With the continuous expansion of wind turbine scale and the increasing complexity of operating environments, traditional detection methods can no longer meet the demands for efficient and accurate fault detection.

[0003] With the advancement of industrial intelligence and the development of sensing technology, blade fault detection technology is gradually evolving towards automation and intelligence. Currently, it can be broadly divided into two categories: one is physical model-based detection methods, which predict faults by establishing physical models of the blades. However, due to the complexity and uncertainty of the wind power equipment's operating environment, physical models often fail to accurately reflect the actual operating state, resulting in insufficient detection accuracy. The other category is data-driven detection methods, particularly those utilizing machine learning techniques for fault detection. However, most existing data-driven methods rely on supervised learning, requiring a large amount of labeled data, which is often difficult to obtain in practical applications. Furthermore, traditional machine learning models such as Support Vector Machines (SVM) and Random Forests, while capable of extracting data features to some extent, struggle to capture the latent representations and complex patterns in audio signals, leading to limited detection effectiveness.

[0004] In recent years, unsupervised learning techniques have shown great potential in the field of fault detection, especially deep learning-based unsupervised methods such as autoencoders and generative adversarial networks (GANs), which can effectively extract latent representations from data without relying on labeled data. However, although deep learning-based unsupervised methods can effectively extract latent representations from data in general scenarios, their performance in the specific application of wind turbine blade fault detection still faces significant challenges. This is mainly due to the unique complexity of the wind power scenario: the wind turbine operating environment is filled with high-intensity, wide-bandwidth environmental noise (such as wind noise, gearbox and generator noise), and the early acoustic signals generated by blade faults (such as cracks, surface damage, or foreign object impacts) are often extremely weak and transient, and their energy and characteristic frequency bands are easily submerged or confused by strong background noise. In addition, the rotational speed of wind turbine blades is constantly changing due to wind speed, resulting in significant non-stationarity and time-varying characteristics in their acoustic features (such as transmission frequencies and harmonics). Existing unsupervised methods often struggle to robustly capture noise and weak fault signals when dealing with highly complex audio signals that are subject to strong noise interference and highly dynamic characteristics. They are also unable to adaptively capture and characterize these transient fault features that drift with the operating state and have varied morphologies, thus limiting their reliability and accuracy in actual wind turbine blade fault detection.

[0005] Therefore, developing a detection method that does not require fault samples, can adapt to strong noise and variable operating conditions, and is sensitive to early faults has important theoretical and engineering significance. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide an unsupervised wind turbine blade fault detection method based on a phase-aware parallel attention mechanism, which solves the problems of existing wind turbine blade fault detection methods being highly dependent on labeled data, having insufficient generalization ability under strong noise and variable operating conditions, and being unable to robustly capture weak transient fault signals and dynamic change characteristics.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0008] The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism uses a trained fault detection model for detection. The fault detection model is a deep adversarial autoencoder architecture, including an encoder, decoder, auxiliary encoder, discriminator, and anomaly scoring module. The encoder, decoder, and auxiliary encoder all have built-in phase-aware parallel attention modules. The phase-aware parallel attention module has high-frequency and low-frequency branches. The high-frequency branch directly receives input features, generates attention weights based on the phase information in the input features, and applies them to the input features to capture transient features in the time dimension.

[0009] The low-frequency branch reduces the dimensionality of the input features through an average pooling layer to obtain dimensionality-reduced features. Attention weights are generated based on the phase information in the input features and applied to the dimensionality-reduced features to capture global features in the frequency dimension.

[0010] The transient features in the time dimension of the high-frequency branch output and the global features in the frequency dimension of the low-frequency branch output are fused to obtain the enhanced features output by the phase-aware parallel attention module.

[0011] The detection method includes the following steps:

[0012] S1. Collect audio signals from the wind turbine blades during operation;

[0013] S2. The acquired audio signal is preprocessed, and the improved short-time Fourier transform is used to extract the dual-channel time-frequency features containing the amplitude spectrum and the phase spectrum.

[0014] S3. After the dual-channel time-frequency features are initially processed by the convolutional unit in the encoder, they are input into its built-in phase-aware parallel attention module for further processing to generate a primary latent representation.

[0015] S4. The primary latent representation is input into the decoder, first processed by its built-in phase-aware parallel attention module, and then processed by the deconvolution unit in the decoder to reconstruct the reconstructed representation consistent with the dual-channel time-frequency feature dimension.

[0016] S5. The reconstructed representation is input into the auxiliary encoder, first processed by its convolutional unit, and then input into its built-in phase-aware parallel attention module for further processing to generate secondary latent representations.

[0017] S6. The anomaly scoring module calculates anomaly scores based on the deviation between primary and secondary latent representations and the deviation between the reconstructed representation and the dual-channel time-frequency features.

[0018] S7. Dynamically set the decision threshold based on the extreme value theory. If the abnormal score is greater than the threshold, it is determined that the wind turbine blades are faulty.

[0019] Furthermore, both the high-frequency and low-frequency branches of the phase-aware parallel attention module are equipped with phase-aware multi-head self-attention modules. Each phase-aware multi-head self-attention module includes multiple parallel attention heads, and each attention head includes three parallel branches: two phase extraction branches with identical structures and one value vector generation branch. The process of attention calculation performed by the phase-aware multi-head self-attention module includes:

[0020] (1) Each attention head calculates the weighted features in the following way:

[0021] Two phase extraction branches are used to perform two-dimensional fast Fourier transform on the input features, and the phase information is obtained after convolution processing.

[0022] The value vector is obtained by performing convolution and depthwise convolution on the input features through the value vector generation branch.

[0023] The phase information generated by the two phase extraction branches is used as the query vector and key vector, respectively, and the weight matrix is ​​calculated.

[0024] Apply the weight matrix to the value vector to generate weighted features;

[0025] (2) The weighted features of all attention heads are spliced ​​and projected to obtain the final feature representation, which is used as the output of the phase-aware multi-head self-attention module.

[0026] Furthermore, in step S2, the acquired audio signal is preprocessed, and an improved short-time Fourier transform is used to extract dual-channel time-frequency features containing amplitude and phase spectra, specifically including:

[0027] The acquired audio signal is pre-emphasized and then divided into frames with a frame length of 40ms.

[0028] Apply a composite window function to each audio frame after framing. ,in, It is a Gaussian window function. The fundamental frequency of the wind turbine blades is estimated in real time; the window length is dynamically calculated as follows: ,in, The sampling frequency of the audio signal;

[0029] Perform a Fast Fourier Transform on the windowed audio frame to obtain the complex spectrum;

[0030] Separating the amplitude spectrum from the complex spectrum and phase spectrum and output dual-channel features ,in, This indicates the number of dimensions of the data, i.e., the channel dimension is 2, the time dimension is T, and the frequency dimension is F.

[0031] Furthermore, the fault detection model is trained offline in an unsupervised manner, and the training process includes:

[0032] a. Collect audio signals from wind turbine blades during normal operation in the past to construct a training dataset containing only non-fault samples;

[0033] b. The audio signals in the training dataset are preprocessed, and the improved short-time Fourier transform is used to extract dual-channel time-frequency features containing amplitude spectrum and phase spectrum;

[0034] c. Based on the dual-channel time-frequency characteristics, the encoder is used to obtain the primary latent representation;

[0035] d. Based on the primary latent representation, the reconstructed representation is obtained using a decoder;

[0036] e. Based on the reconstructed representation, secondary latent representations are obtained using an auxiliary encoder;

[0037] f. Using the dual-channel time-frequency features and reconstructed representation as inputs to the discriminator, calculate the adversarial loss based on the output of the discriminator;

[0038] g. Based on the adversarial loss, calculate the composite loss function by jointly considering the reconstruction error loss, the potential representation consistency loss, and the phase consistency loss;

[0039] h. Based on the composite loss function, perform backward gradient updates of the model to optimize network parameters;

[0040] i. Repeat step bh until the model parameters converge to obtain a trained fault detection model.

[0041] Furthermore, in step g, the composite loss function is expressed as:

[0042] ;

[0043] in, This represents the total loss of the model; , , They are preset , , The weights;

[0044] To compensate for reconstruction error loss, ;

[0045] in, It has dual-channel time-frequency characteristics. To reconstruct the representation;

[0046] For potential representation consistency loss, ;

[0047] in, This is the initial potential characterization. For secondary potential representation;

[0048] To combat the losses, ;

[0049] in, The implicit function representing the discriminator. Represents the Sigmoid function;

[0050] For phase consistency loss, ;

[0051] in, For time attention weights, For frequency attention weights, Indicates gradient calculation, Represents the tensor product. This represents total variation computation. The coefficient is 0.5.

[0052] Furthermore, in step S6, the anomaly scoring module calculates the anomaly score based on the deviation between the primary latent representation and the secondary latent representation, as well as the deviation between the reconstructed representation and the dual-channel time-frequency features. The calculation methods include:

[0053] ;

[0054] in, For abnormal scoring, For dynamic weights, For feature distance scoring, Scoring is given for reconstruction error;

[0055] ;

[0056] ;

[0057] in, For phase unwinding operation; , These are the amplitude spectrum and phase spectrum of the input audio signal, respectively; , These are the reconstructed characterization and its phase spectrum, respectively; The preset parameter is 0.3; This represents the inverse of the covariance matrix.

[0058] Furthermore, dynamic weights The calculations are based on environmental noise levels, specifically including:

[0059] ;

[0060] in, For the estimated signal-to-noise ratio, To adjust the parameters, the noise spectrum Estimated by silent segments It is the squared Frobenius norm of the matrix.

[0061] Furthermore, in step S7, the decision threshold is dynamically set according to the extreme value theory, including:

[0062] First, obtain the anomaly score within the sliding window. mean and standard deviation ;

[0063] Then according to the formula Calculate the extreme value coefficients, where, Use the preset value of 0.99;

[0064] Finally, calculate the decision threshold. .

[0065] Furthermore, the sliding window is set to an integer multiple of the complete rotation cycle of the wind turbine blade.

[0066] Furthermore, in step S1, the method for acquiring the audio signal of the wind turbine blades during operation includes:

[0067] By deploying a microphone array on the wind turbine blades to be inspected, audio signals of the wind turbine blades during operation are collected, and the collected audio signals are transmitted wirelessly to a remote smart terminal, which is equipped with a trained fault detection model.

[0068] The beneficial effects of this invention are:

[0069] (1) The fault detection model adopted in this invention is based on the encoder-decoder-encoder architecture, and innovatively introduces a phase-aware parallel attention module in the encoder and decoder. This module, through the parallel dual-branch architecture (high frequency / low frequency) and the design of explicitly integrating phase information into attention calculation, can simultaneously capture the weak transient abnormal features generated by blade faults from the time dimension and identify the characteristic frequency band drift caused by speed fluctuations from the frequency dimension under strong background noise (such as wind noise and mechanical noise). This allows for more robust modeling of the potential fault characterization of wind turbine blades in complex dynamic operating environments, thereby improving the reliability and accuracy of wind turbine blade fault detection.

[0070] (2) The model in this invention adopts an offline unsupervised training method, which only requires samples under normal working conditions to complete the model training, thus eliminating the dependence on scarce fault labeling data. At the same time, in the feature learning stage, based on the introduced discriminator, adversarial training is carried out with the generator network. Combined with the phase consistency loss function to constrain the smoothness of attention weights in the time-frequency dimension, it is ensured that the learned features conform to the physical laws of sound wave propagation, thereby optimizing the spatial distribution of audio features of wind turbine blades under normal conditions, improving the sensitivity to abnormal features, and enabling the model to maintain stable performance under strong noise and variable operating conditions.

[0071] (3) When the trained model is applied to actual fault detection, the present invention adopts a multi-scale anomaly scoring algorithm to construct a composite detection index by combining feature space distance and reconstruction error, and dynamically sets the decision threshold by combining extreme value theory, thereby achieving high-precision fault warning and significantly improving the fault detection rate. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the fault detection model architecture in an embodiment of the present invention.

[0073] Figure 2 This is a schematic diagram of an application scenario in an embodiment of the present invention.

[0074] Figure 3 This is a schematic diagram of the generative adversarial network structure in an embodiment of the present invention.

[0075] Figure 4 This is a schematic diagram of the phase-aware parallel attention mechanism in an embodiment of the present invention.

[0076] Figure 5 This is a schematic diagram of the phase-aware multi-head self-attention mechanism in an embodiment of the present invention. Detailed Implementation

[0077] This invention aims to provide an unsupervised wind turbine blade fault detection method based on a phase-aware parallel attention mechanism. It addresses the problems of existing wind turbine blade fault detection methods, such as strong dependence on labeled data, insufficient generalization ability under strong noise and variable operating conditions, and difficulty in robustly capturing weak transient fault signals and dynamic characteristics. The core idea is to address the characteristics of weak, transient, easily submerged, and non-stationary time-varying acoustic signals of wind turbine blade faults. This is achieved through a phase-aware mechanism to enhance sensitivity to subtle waveform distortions, dynamic STFT to adapt to speed changes, and multi-scale anomaly scoring fusion (feature offset and signal distortion). These core designs significantly improve the detection accuracy, robustness, and generalization ability of early blade faults (such as microcracks and foreign object impacts) under conditions without labeled data, providing more reliable technical support for intelligent operation and maintenance of wind turbines.

[0078] More specific means employed by this invention to achieve the aforementioned core ideas include, but are not limited to:

[0079] (1) The detection model is based on the encoder-decoder-encoder architecture and innovatively introduces a phase-aware parallel attention module in the encoder and decoder. This module, through the parallel dual-branch architecture (high frequency / low frequency) and the design of explicitly integrating phase information into attention calculation, can simultaneously capture the weak transient abnormal features (such as crack impact waveform) caused by blade failure from the time dimension and identify the characteristic frequency band drift and structural relationship changes caused by speed fluctuations from the frequency dimension under strong background noise (such as wind noise and mechanical noise), thereby more robustly modeling the potential fault characterization of wind turbine blades in complex dynamic operating environments.

[0080] (2) When preprocessing the audio signal collected during the operation of the wind turbine blade, an improved short-time Fourier transform is used. This process uses a composite window function to window the audio signal, effectively focusing and enhancing the weak fault signal related to blade rotation, and simultaneously outputting both amplitude and phase spectrum information, providing richer and more targeted time-frequency feature representations for subsequent analysis. Furthermore, during the windowing process, the window length is dynamically adjusted according to the real-time estimated blade rotation frequency to ensure that the complete cycle features of blade rotation can be captured in each analysis.

[0081] (3) The model adopts an offline unsupervised training method, which only requires normal state samples to complete the model training, thus eliminating the dependence on scarce fault labeling data. At the same time, in the feature learning stage, based on the introduced discriminator, it conducts adversarial training with the generator network, and combines the phase consistency loss function to constrain the smoothness of the attention weights in the time-frequency dimension, ensuring that the learned features conform to the physical laws of sound wave propagation, thereby optimizing the spatial distribution of audio features of wind turbine blades under normal conditions, improving the sensitivity to abnormal features, and enabling the model to maintain stable performance under strong noise and variable operating conditions.

[0082] (4) When performing fault detection, a multi-scale anomaly scoring algorithm is used to construct a composite detection index by combining feature space distance and reconstruction error. The feature space distance score reflects the state deviation of the potential space (systematic anomaly pattern), while the reconstruction error score directly captures the physical distortion of the original signal (including amplitude error of energy anomaly and phase error of waveform distortion), thereby revealing the fault from different dimensions. The decision threshold is dynamically set in combination with extreme value theory to make fault judgment. The setting of this dynamic decision threshold fully considers the uncertainty in the operation of wind power equipment, thereby realizing high-precision fault warning and significantly improving the fault detection rate.

[0083] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0084] The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism provided in this embodiment uses a trained fault detection model to achieve detection.

[0085] The fault detection model is a deep adversarial autoencoder architecture, see [link / reference]. Figure 1 It includes encoder E1, decoder D1, auxiliary encoder E2, discriminator, and anomaly scoring module; during the training phase, encoder E1, decoder D1, auxiliary encoder E2, and discriminator constitute an adversarial network, such as... Figure 3 As shown, feature extraction is performed on training samples (audio data of wind turbine blades during normal operation in the past) to obtain dual-channel time-frequency features containing amplitude spectrum and phase spectrum. These features are processed by encoder E1 to obtain primary latent representation. The primary latent representation is input to decoder D1 to obtain reconstructed representation. The reconstructed representation is input to auxiliary encoder E2 to obtain secondary latent representation. Based on the dual-channel time-frequency features and reconstructed representation as input to the discriminator, the discriminator distinguishes the authenticity of the original spectrum from the reconstructed spectrum, thereby calculating the adversarial loss. Then, the reconstruction error loss, latent representation consistency loss and phase consistency loss are combined to calculate the composite loss function, and the model parameters are updated by back gradient, thereby iteratively optimizing the model parameters.

[0086] exist Figure 3 In the schematic structure, the input feature X is a preprocessed dual-channel time-frequency feature, including the amplitude spectrum. and phase spectrum The encoder E1 consists of multiple convolutional modules (convolutional module 1, convolutional module 2, and convolutional module 3) connected sequentially, along with a phase-aware parallel attention module. The convolutional modules progressively compress the feature dimensions through convolution operations, while the phase-aware parallel attention module enhances the temporal and frequency-critical information of the convolutional features, ultimately outputting a primary latent representation. The decoder D1 adopts a structure symmetrical to the encoder E1, containing a phase-aware parallel attention module and multiple deconvolution modules (deconvolution module 1, deconvolution module 2, and deconvolution module 3); the phase-aware parallel attention module processes the primary latent representations. After enhancing the key time-frequency information, the dimensions are gradually restored through a deconvolution module to obtain the reconstructed representation. The auxiliary encoder E2 has a similar structure to the encoder E1, consisting of multiple convolutional modules (convolutional module 1, convolutional module 2, and convolutional module 3) and a phase-aware parallel attention module, which receives the reconstructed representation output by the decoder D1. The feature dimensions are gradually compressed through convolution operations, and then a phase-aware parallel attention module is used to enhance the temporal and frequency key information of the convolutional features to generate secondary latent representations. Discriminator Used to distinguish input features With reconstructed representation The authenticity of the input is determined by simultaneously connecting the original input features. and reconstructed representation The output is used to calculate the adversarial loss. Furthermore, skip connections exist between the corresponding convolutional and deconvolutional modules of encoder E1 and decoder D1, directly passing low-level feature information to help improve reconstruction accuracy. The entire network utilizes primary latent representations... With secondary potential representation Consistency constraints, input features With reconstructed representation Error constraints and adversarial constraints of the discriminator enable accurate learning of the distribution of features in the normal state.

[0087] The structure of the phase-aware parallel attention module is as follows: Figure 4 As shown, it adopts a parallel dual-branch architecture: the high-frequency branch directly inputs the dual-channel time-frequency features into the phase-aware multi-head self-attention module to capture transient anomaly features in the time dimension; the low-frequency branch extracts global anomaly patterns in the frequency dimension by performing average pooling dimensionality reduction on the dual-channel time-frequency features and then extracting them through the same attention module. Finally, the output features of the two branches are spliced ​​and projected to form an enhanced representation that combines high-frequency details and low-frequency structures.

[0088] More specifically, the process of the phase-aware parallel attention module is as follows:

[0089] High-frequency branches directly incorporate input features The high-frequency branch attention features are obtained by feeding them into a phase-aware multi-head self-attention module. Then to Perform splicing and projection (weight matrix is) This operation generates high-frequency branch output features. The low-frequency branch first performs average pooling on the input features to obtain pooled features. Then, by undergoing phase-aware multi-head self-attention again, low-frequency branch attention features are obtained. Then to Perform splicing and projection (weight matrix) ), generate low-frequency branch output features Finally, the output features of the two branches are... and The splicing features are obtained by splicing. And through the final projection layer (weight matrix) The output features of the fusion module are obtained. .

[0090] The structure of the aforementioned phase-aware multi-head self-attention module is as follows: Figure 5 As shown, this module is the core component of the high-frequency and low-frequency branches in the phase-aware parallel attention module. The input is the feature matrix from the high-frequency or low-frequency branch. It is responsible for integrating phase information into the attention weight calculation. It contains multiple attention heads, and each attention head contains three parallel processing branches.

[0091] Phase extraction branch 1: After performing a 2D Fast Fourier Transform on the input, it sequentially undergoes 1×1 convolution (compressing channel dimensions), depthwise convolution (extracting local features), and shape reshaping (adjusting matrix dimensions) operations to output the phase information Φ used to generate the query vector. q .

[0092] Phase extraction branch 2: It has the same structure as phase extraction branch 1, but with different learnable projective weights. After performing a 2D Fast Fourier Transform on the input, it sequentially undergoes 1×1 convolution (compressing channel dimensions), depthwise convolution (extracting local features), and shape reshaping (adjusting matrix dimensions), outputting phase information Φ used to generate the key vector. k .

[0093] Value vector generation branch: The input is directly processed by 1×1 convolution, depthwise convolution and shape reshaping operations to output a value vector V, which is used to carry the feature content.

[0094] Attention calculation: Φ q and Φ k The phase correlation matrix is ​​obtained by element-wise multiplication. After normalization by the Softmax layer, attention weights are generated. These weights are then multiplied element-wise with the value vector V to finally output the weighted feature vector.

[0095] This module explicitly extracts phase information and incorporates an attention mechanism, enabling the model to accurately capture fault-related phase distortions in time-frequency features and enhance the ability to identify subtle fault signals.

[0096] During the inference phase, feature extraction is performed on the collected audio data of the wind turbine blades to be tested, obtaining dual-channel time-frequency features including amplitude and phase spectra. These features are processed by encoder E1 to obtain a primary latent representation. The primary latent representation is input to decoder D1 to obtain a reconstructed representation. The reconstructed representation is input to auxiliary encoder E2 to obtain a secondary latent representation. The anomaly scoring module calculates an anomaly score based on the deviation between the primary and secondary latent representations and the deviation between the reconstructed representation and the dual-channel time-frequency features. This anomaly score is compared with a decision threshold dynamically set according to extreme value theory. If the anomaly score is greater than the threshold, the wind turbine blades are determined to have a fault.

[0097] One application scenario of the above fault detection model is as follows: Figure 2As shown, high-precision sound acquisition devices are deployed on wind turbine blades to collect audio data during blade rotation. In one exemplary implementation, a high-precision microphone array is deployed on the wind turbine blades to collect audio signals generated during blade operation, with a sampling frequency ≥44.1kHz. Time-delay beamforming is used to enhance the target sound pressure. The collected audio data is then transmitted wirelessly to a remote smart terminal. The smart terminal is equipped with a trained fault detection model. Upon receiving the uploaded audio data, the model is used to process the audio data to detect whether a fault has occurred on the current wind turbine blade.

[0098] In one exemplary implementation, the fault detection scheme for wind turbine blades using a fault detection model by the intelligent terminal includes the following processing steps:

[0099] S1. Audio signal preprocessing, extracting dual-channel time-frequency features:

[0100] In this step, the acquired audio signal is preprocessed, and an improved short-time Fourier transform is used to extract dual-channel time-frequency features containing amplitude and phase spectra.

[0101] Specifically, it includes:

[0102] The acquired audio signal is pre-emphasized and then divided into frames with a frame length of 40ms.

[0103] Apply a composite window function to each audio frame after framing. ,in, It is a Gaussian window function. The fundamental frequency of the wind turbine blades is estimated in real time; the window length is dynamically calculated as follows: ,in, The sampling frequency of the audio signal;

[0104] Perform a Fast Fourier Transform on the windowed audio frame to obtain the complex spectrum;

[0105] Separating the amplitude spectrum from the complex spectrum and phase spectrum and output dual-channel features ,in, This indicates the number of dimensions of the data, i.e., the channel dimension is 2, the time dimension is T, and the frequency dimension is F.

[0106] The key difference between the improved short-time Fourier transform (SFT) used in this embodiment and the ordinary SFT lies in its highly adaptable characteristic to changes in wind turbine blade rotation speed. Conventional methods typically use a fixed-length time window for analysis, while this method dynamically adjusts the window length based on the real-time estimated blade rotation frequency, ensuring that each analysis captures the complete cyclic characteristics of blade rotation. Simultaneously, it employs a special composite window design to effectively focus on and enhance weak fault signals related to blade rotation, and simultaneously outputs both amplitude and phase spectrum information, providing richer and more targeted time-frequency feature representations for subsequent analysis. The obtained dual-channel features are phase-sensitive, effectively suppressing spectral leakage caused by rotation speed fluctuations, while explicitly preserving the complete physical information of the amplitude and phase spectra, providing high-fidelity input for subsequent processing.

[0107] S2. The encoder processes the dual-channel time-frequency features to generate a primary latent representation:

[0108] In this step, the dual-channel time-frequency features are initially processed by the convolutional units in the encoder, and then input into its built-in phase-aware parallel attention module for further processing to generate a primary latent representation. .

[0109] S3. The decoder processes the primary latent representation to obtain the reconstructed representation:

[0110] In this step, the primary latent representation The input to the decoder is first processed by its built-in phase-aware parallel attention module, and then processed by the deconvolution unit in the decoder to reconstruct the dual-channel time-frequency features. Dimensionally Consistent Reconstruction Representation .

[0111] S4. The auxiliary encoder processes the reconstructed representation to generate secondary latent representations:

[0112] In this step, the characterization is reconstructed. The input to the auxiliary encoder undergoes preliminary processing via its convolutional units, followed by further processing via its built-in phase-aware parallel attention module to generate secondary latent representations. .

[0113] S5. Calculate the anomaly score using the anomaly scoring module:

[0114] In this step, the anomaly scoring module is based on primary latent representations. With secondary potential representation Bias and Reconstructed Characterization With dual-channel time-frequency characteristics The deviation is used to calculate the anomaly score.

[0115] In this embodiment, a multi-scale anomaly scoring algorithm is used to calculate the anomaly score, mainly considering the feature distance score and the reconstruction error score. The feature distance score calculates the primary latent representation. With secondary potential representation The standardized Mahalanobis distance reflects the state shift of the potential space (systematic anomalous patterns); the reconstruction error score, which combines amplitude error and unwinding phase error, can directly capture the physical distortion of the original signal (including amplitude error of energy anomalies and phase error of waveform distortion); finally, the two scores are fused by dynamic weighting to achieve a comprehensive revelation of the fault from different dimensions.

[0116] Feature distance score The calculation method is as follows:

[0117] ;

[0118] in, This represents the inverse of the covariance matrix.

[0119] Reconstruction error score The calculation method is as follows:

[0120] ;

[0121] in, For phase unwinding operation; These are the amplitude and phase spectra of the input audio. and These are the reconstructed characterization and its phase spectrum. , is a preset parameter.

[0122] Final rating The calculation method is as follows:

[0123] ;

[0124] in, The weights are dynamic and adaptively adjusted based on the environmental signal-to-noise ratio. Specific calculation methods include:

[0125] ;

[0126] in, The estimated signal-to-noise ratio; To adjust parameters; noise spectrum Estimated by the silent segment; It is the square of the Frobenius norm of the matrix, which is the sum of the squares of the absolute values ​​of all elements of the matrix.

[0127] In this embodiment, the reliability of both can be balanced under different signal-to-noise ratios by using environmentally adaptive dynamic weight fusion, that is, focusing on reconstructing details when the signal-to-noise ratio is high, and relying on feature distribution when the signal-to-noise ratio is low.

[0128] S6. Determine whether wind turbine blades are faulty based on dynamic decision thresholds:

[0129] In this step, the decision threshold is dynamically set based on extreme value theory, and the anomaly score calculated in step S5 is used to... If the anomaly score is greater than the decision threshold, the wind turbine blades are determined to be faulty.

[0130] In this embodiment, the decision threshold is dynamically set using extreme value theory, and the specific calculation method is as follows:

[0131] First, obtain the composite anomaly score within the sliding window (the data time range used for dynamic statistical anomaly scoring during real-time system monitoring, usually set to an integer multiple of the complete rotation cycle of the wind turbine blade). mean and standard deviation .

[0132] Then, according to the formula Calculate the extreme value coefficients (in Use the preset value of 0.99.

[0133] Finally, a dynamic decision threshold is generated. .

[0134] When real-time calculation of anomaly scores The value exceeds A fault alarm will be triggered immediately.

[0135] By dynamically setting decision thresholds using extreme value theory, we can adapt to fluctuations in operating conditions and significantly improve the robustness and accuracy of detection.

[0136] The training process of the fault detection model is described below:

[0137] The fault detection model in this embodiment is trained offline without supervision. The training process includes:

[0138] a. Collect audio signals from wind turbine blades during normal operation in the past to construct a training dataset containing only non-fault samples;

[0139] b. The audio signals in the training dataset are preprocessed, and the improved short-time Fourier transform is used to extract dual-channel time-frequency features containing amplitude spectrum and phase spectrum;

[0140] c. Based on the dual-channel time-frequency characteristics, the encoder is used to obtain the primary latent representation;

[0141] d. Based on the primary latent representation, the reconstructed representation is obtained using a decoder;

[0142] e. Based on the reconstructed representation, secondary latent representations are obtained using an auxiliary encoder;

[0143] f. Using dual-channel time-frequency features and reconstructed representations as inputs to the discriminator, the adversarial loss is calculated based on the discriminator's output;

[0144] g. Based on the adversarial loss, calculate the composite loss function by jointly considering the reconstruction error loss, the potential representation consistency loss, and the phase consistency loss;

[0145] h. Based on the composite loss function, perform backward gradient updates of the model to optimize network parameters;

[0146] i. Repeat step bh until the model parameters converge to obtain a trained fault detection model.

[0147] Based on the above training method, model training can be completed using only normal state samples, eliminating the dependence on scarce fault-labeled data. Simultaneously, during the feature learning stage, adversarial training is conducted between the introduced discriminator and the generator network. Combined with the phase consistency loss function to constrain the smoothness of attention weights in the time-frequency dimension, this ensures that the learned features conform to the physical laws of sound wave propagation. This optimizes the spatial distribution of audio features of wind turbine blades under normal conditions, enhances sensitivity to abnormal features, and enables the model to maintain stable performance under strong noise and varying operating conditions.

[0148] Specifically, this embodiment employs joint optimization of four types of loss functions during the model training phase: reconstruction error loss. Potential representation consistency loss Combating losses Phase consistency loss ;

[0149] Among them, reconstruction error loss By directly constraining the reconstructed representation generated by the decoder to maintain consistency with the original input features (including amplitude and phase spectra), the model can accurately learn the detailed patterns of time-frequency features under normal operating conditions (such as amplitude distribution patterns and phase continuity). This constraint forces the decoder to faithfully reproduce the key information of the input features, reducing reconstruction bias under normal conditions. The calculation is as follows:

[0150] ;

[0151] in, It has dual-channel time-frequency characteristics. To reconstruct the representation.

[0152] Potential representation consistency loss By ensuring the consistency in distribution between the primary latent representation generated by the constraint encoder and the secondary latent representation generated by the auxiliary encoder, it is guaranteed that the features reconstructed by the decoder can still be accurately mapped back to the original latent feature space after passing through the auxiliary encoder. This constraint strengthens the model's stable representation ability of features under normal operating conditions, reduces the loss or distortion of feature information during reconstruction, and enables the model to more accurately learn the feature distribution patterns of normal states. Therefore, during the detection phase, the deviation between the two representations can be used to effectively identify abnormal states. The calculation is as follows:

[0153] ;

[0154] in, This is the initial potential characterization. This is a secondary potential representation.

[0155] Combat losses By using a discriminator to distinguish between genuine and fake features in the original dual-channel time-frequency features and decoder-reconstructed features, the generator is driven to produce reconstructed features that more closely resemble the actual physical characteristics. This loss enhances the realism of the reconstructed features in the time-frequency domain (such as the reasonableness of the amplitude spectrum distribution and the physical continuity of the phase spectrum), preventing the model from simply replicating features for reconstruction. Instead, it learns the essential laws of features under normal operating conditions, thereby improving the sensitivity to feature anomalies caused by subtle faults. Its calculation is as follows:

[0156] ;

[0157] in, The implicit function representing the discriminator. This represents the Sigmoid function.

[0158] Phase consistency loss This method combines physical acoustic properties with deep learning attention mechanisms using mathematical formulas to constrain the consistency of the original dual-channel time-frequency features and reconstructed features in the phase spectrum. This includes numerical deviations after phase spectrum unwrapping and deviations in the continuity of the phase derivative with respect to time / frequency. Its core function is to ensure that the phase information of the reconstructed features conforms to physical laws (such as the phase change trend corresponding to the blade rotation fundamental frequency), avoiding modeling deviations in normal states due to phase distortion. Accurate restoration of phase information enhances the model's ability to capture phase abrupt changes caused by faults (such as vibration phase shifts caused by blade cracks), supplementing the lack of amplitude spectrum information. The calculation is as follows:

[0159] ;

[0160] in, For time attention weights, For frequency attention weights, Indicates gradient calculation, Represents the tensor product. This represents total variation computation. The coefficient is 0.5.

[0161] In the above equation, the total variation regularization term This is a smoothness constraint on time-frequency attention. This term affects the weights of time-frequency attention. and frequency attention weight Calculate the gradient separately Then, through tensor product Construct the joint time-frequency gradient matrix, and then calculate its total variation. This makes the attention weights change smoothly in both time and frequency dimensions, avoiding drastic fluctuations caused by noise interference.

[0162] Finally, the composite loss function is expressed as:

[0163] ;

[0164] in, This represents the total loss of the model; , , They are preset , , The weight.

[0165] In one exemplary implementation, the training process is as follows:

[0166] First, a dedicated training dataset is constructed: continuous audio is collected during the normal operation of wind power equipment, and raw waveforms with a duration of no less than 200 hours are obtained through a microphone array.

[0167] The data preprocessing follows the process described above: the original signal is pre-emphasized to enhance high-frequency components, and after being framed at a frame length of 40ms (50% overlap), a dynamic window length STFT transformation is performed using a composite window function to generate a dual-channel time-frequency graph sample set containing amplitude spectrum and phase spectrum, ultimately forming approximately 150,000 training samples.

[0168] Key training control measures include: verifying the reconstructed signal-to-noise ratio after each training round, triggering gradient pruning when it falls below 35dB; and linearly increasing the weight of the phase consistency loss with each training round, reaching a maximum of 40% of the total loss. The Adam optimizer is used for backpropagation, with an initial learning rate of 5×10⁻⁻⁶. 4 Furthermore, the training decays by 30% every 50 rounds, for a total of 300 rounds.

[0169] The entire training process took approximately 36 hours to complete on a single NVIDIA A5000 GPU, consuming only normal operating data and requiring no fault labeling.

[0170] Verification example:

[0171] To verify the accuracy of the present invention in detecting wind turbine blade faults, this verification example will compare the detection accuracy of three different wind turbine blade detection models. The test audio data includes hundreds of audio data collected from different parts of different wind turbine blades under normal or abnormal conditions. The detection models used include the model proposed in this invention, the CNN (convolutional neural network) model and the CRNN (convolutional recurrent neural network) model in the prior art.

[0172] The detection accuracy of the proposed model, the detection accuracy of the CNN model, and the detection accuracy of the CRNN model are shown in Tables 1, 2, and 3, respectively.

[0173] Table 1. Detection accuracy of the model proposed in this invention.

[0174]

[0175] Table 2 Detection Accuracy of CNN Models

[0176]

[0177] Table 3 Detection Accuracy of CRNN Model

[0178]

[0179] In the table above, f1, f2, and f3 represent three different types of wind turbines used onshore, offshore, and high-altitude areas, respectively. Part 1 is the blade root bolt connection area, part 2 is the leading edge wind erosion protection layer, and part 3 is the trailing edge main beam cap area.

[0180] The data shows that the method proposed in this invention achieves average detection accuracies of 85.54%, 87.43%, and 84.35% on devices f1, f2, and f3, respectively, all significantly higher than the average detection accuracies of CNN and CRNN models on these three devices. Furthermore, the method of this invention also achieves higher accuracy than the other two models for the same location. This indicates that the method proposed in this invention has better accuracy and reliability in wind turbine blade fault detection, can more effectively identify blade fault conditions, and provides stronger support for the maintenance and management of wind turbines.

[0181] Therefore, although embodiments of the present invention have been described above, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and variations shall not depart from the protection scope of the present invention.

Claims

1. An unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism, which uses a trained fault detection model for detection, characterized in that... The fault detection model adopts a deep adversarial autoencoder architecture, including an encoder, a decoder, an auxiliary encoder, a discriminator, and an anomaly scoring module. The encoder, decoder, and auxiliary encoder all have a built-in phase-aware parallel attention module. The phase-aware parallel attention module has a high-frequency branch and a low-frequency branch. The high-frequency branch directly receives input features, generates attention weights based on the phase information in the input features, and applies them to the input features to capture transient features in the time dimension. The low-frequency branch reduces the dimensionality of the input features through an average pooling layer to obtain dimensionality-reduced features. Attention weights are generated based on the phase information in the input features and applied to the dimensionality-reduced features to capture global features in the frequency dimension. The transient features in the time dimension of the high-frequency branch output and the global features in the frequency dimension of the low-frequency branch output are fused to obtain the enhanced features output by the phase-aware parallel attention module. The detection method includes the following steps: S1. Collect audio signals from the wind turbine blades during operation; S2. The acquired audio signal is preprocessed, and the improved short-time Fourier transform is used to extract the dual-channel time-frequency features containing the amplitude spectrum and the phase spectrum. S3. After the dual-channel time-frequency features are initially processed by the convolutional unit in the encoder, they are input into its built-in phase-aware parallel attention module for further processing to generate a primary latent representation. S4. The primary latent representation is input into the decoder, first processed by its built-in phase-aware parallel attention module, and then processed by the deconvolution unit in the decoder to reconstruct the reconstructed representation consistent with the dual-channel time-frequency feature dimension. S5. The reconstructed representation is input into the auxiliary encoder, first processed by its convolutional unit, and then input into its built-in phase-aware parallel attention module for further processing to generate secondary latent representations. S6. The anomaly scoring module calculates anomaly scores based on the deviation between primary and secondary latent representations and the deviation between the reconstructed representation and the dual-channel time-frequency features. S7. Dynamically set the decision threshold based on the extreme value theory. If the abnormal score is greater than the threshold, it is determined that the wind turbine blades are faulty.

2. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism as described in claim 1, characterized in that, The phase-aware parallel attention module includes phase-aware multi-head self-attention modules in both its high-frequency and low-frequency branches. Each phase-aware multi-head self-attention module comprises multiple parallel attention heads, and each attention head includes three parallel branches: two phase extraction branches with identical structures and one value vector generation branch. The attention calculation process performed by the phase-aware multi-head self-attention module includes: (1) Each attention head calculates the weighted features in the following way: Two phase extraction branches are used to perform two-dimensional fast Fourier transform on the input features, and the phase information is obtained after convolution processing. The value vector is obtained by performing convolution and depthwise convolution on the input features through the value vector generation branch. The phase information generated by the two phase extraction branches is used as the query vector and key vector, respectively, and the weight matrix is ​​calculated. Apply the weight matrix to the value vector to generate weighted features; (2) The weighted features of all attention heads are spliced ​​and projected to obtain the final feature representation, which is used as the output of the phase-aware multi-head self-attention module.

3. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism as described in claim 1, characterized in that, In step S2, the acquired audio signal is preprocessed, and an improved short-time Fourier transform is used to extract dual-channel time-frequency features containing amplitude and phase spectra. Specifically, this includes: The acquired audio signal is pre-emphasized and then divided into frames with a frame length of 40ms. Apply a composite window function to each audio frame after framing. ,in, It is a Gaussian window function. The fundamental frequency of the wind turbine blades is estimated in real time; the window length is dynamically calculated as follows: ,in, The sampling frequency of the audio signal; Perform a Fast Fourier Transform on the windowed audio frame to obtain the complex spectrum; Separating the amplitude spectrum from the complex spectrum and phase spectrum and output dual-channel features ,in, This indicates the number of dimensions of the data, i.e., the channel dimension is 2, the time dimension is T, and the frequency dimension is F.

4. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism as described in claim 1, characterized in that, The fault detection model is trained offline in an unsupervised manner. The training process includes: a. Collect audio signals from wind turbine blades during normal operation in the past to construct a training dataset containing only non-fault samples; b. The audio signals in the training dataset are preprocessed, and the improved short-time Fourier transform is used to extract dual-channel time-frequency features containing amplitude spectrum and phase spectrum; c. Based on the dual-channel time-frequency characteristics, the encoder is used to obtain the primary latent representation; d. Based on the primary latent representation, the reconstructed representation is obtained using a decoder; e. Based on the reconstructed representation, secondary latent representations are obtained using an auxiliary encoder; f. Using the dual-channel time-frequency features and reconstructed representation as inputs to the discriminator, calculate the adversarial loss based on the output of the discriminator; g. Based on the adversarial loss, calculate the composite loss function by jointly considering the reconstruction error loss, the potential representation consistency loss, and the phase consistency loss; h. Based on the composite loss function, perform backward gradient updates of the model to optimize network parameters; i. Repeat step bh until the model parameters converge to obtain a trained fault detection model.

5. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism as described in claim 4, characterized in that, In step g, the composite loss function is expressed as: ； in, This represents the total loss of the model; , , They are preset , , The weights; To compensate for reconstruction error loss, ; in, It has dual-channel time-frequency characteristics. To reconstruct the representation; For potential representation consistency loss, ; in, This is the initial potential characterization. For secondary potential representation; To combat the losses, ; in, The implicit function representing the discriminator. Represents the Sigmoid function; For phase consistency loss, ; in, For time attention weights, For frequency attention weights, Indicates gradient calculation, Represents the tensor product. This represents total variation computation. The coefficient is 0.

5.

6. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism as described in claim 5, characterized in that, In step S6, the anomaly scoring module calculates the anomaly score based on the deviation between the primary latent representation and the secondary latent representation, as well as the deviation between the reconstructed representation and the dual-channel time-frequency features. The calculation methods include: ； in, For abnormal scoring, For dynamic weights, For feature distance scoring, Scoring is given for reconstruction error; ； ； in, For phase unwinding operation; , These are the amplitude spectrum and phase spectrum of the input audio signal, respectively; , These are the reconstructed characterization and its phase spectrum, respectively. The preset parameter is 0.3; This represents the inverse of the covariance matrix.

7. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism as described in claim 6, characterized in that, Dynamic weights The calculations are based on environmental noise levels, specifically including: ； in, For the estimated signal-to-noise ratio, To adjust the parameters, the noise spectrum Estimated by silent segments It is the squared Frobenius norm of the matrix.

8. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism as described in claim 7, characterized in that, In step S7, the decision threshold is dynamically set according to the extreme value theory, including: First, obtain the anomaly score within the sliding window. mean and standard deviation ; Then according to the formula Calculate the extreme value coefficients ,in, Use the preset value of 0.99; Finally, calculate the decision threshold. .

9. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism as described in claim 8, characterized in that, The sliding window is set to an integer multiple of the complete rotation cycle of the wind turbine blade.

10. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism as described in any one of claims 1-9, characterized in that, In step S1, the methods for acquiring the audio signal of the wind turbine blades during operation include: By deploying a microphone array on the wind turbine blades to be inspected, audio signals of the wind turbine blades during operation are collected, and the collected audio signals are transmitted wirelessly to a remote smart terminal, which is equipped with a trained fault detection model.

Images