Unsupervised wind power equipment blade fault detection method based on phase perception parallel attention mechanism

Abstract

The invention relates to a wind power equipment blade fault detection technology, discloses an unsupervised wind power equipment blade fault detection method based on a phase perception parallel attention mechanism, and solves the problems that an existing wind power equipment blade fault detection method is high in dependence on labeled data, insufficient in generalization ability under strong noise and variable working conditions and high in fault detection efficiency. And a weak transient fault signal and a dynamic change characteristic are difficult to capture robustly. According to the scheme of the invention, the method comprises the steps: collecting a blade operation audio signal, and extracting a dual-channel time-frequency feature containing an amplitude spectrum and a phase spectrum through improved short-time Fourier transform; a deep adversarial auto-encoder is constructed by using an encoder containing a phase perception parallel attention module, a decoder and an auxiliary encoder, and normal working condition feature distribution is learned by reconstructing an error loss, potential representation consistency loss, adversarial loss and phase consistency loss optimization model during off-line training; in the reasoning stage, the fault is judged based on the feature distance score and the reconstruction error score.

Description

TECHNICAL FIELD

[0001] The present application relates to wind turbine blade fault detection technology, and in particular to an unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism. BACKGROUND

[0002] Wind turbine blade fault detection is a key link in wind power operation, directly affecting the operation efficiency and safety of wind farms. Traditional blade fault detection methods mainly rely on periodic manual inspection and sensor monitoring, which not only has high cost, but also is difficult to achieve real-time and comprehensive fault detection. With the continuous expansion of wind power equipment scale and the increasing complexity of operating environment, traditional detection methods have been unable to meet the demand for efficient and accurate fault detection.

[0003] With the advancement of industrial intelligence and the development of sensing technology, blade fault detection technology has gradually evolved towards automation and intelligence. Currently, it can be roughly divided into two categories: one is the detection method based on physical model, which predicts faults by establishing a physical model of the blade. However, due to the complexity and uncertainty of the wind power equipment operating environment, the physical model often fails to accurately reflect the actual operating state, resulting in insufficient detection accuracy. The other is the data-driven detection method, especially the use of machine learning technology for fault detection. However, existing data-driven methods mostly rely on supervised learning, which requires a large amount of labeled data, which is often difficult to obtain in practical applications. In addition, traditional machine learning models such as support vector machine (SVM) and random forest, although they can extract data features to some extent, they are difficult to capture the potential representation and complex patterns in audio signals, resulting in limited detection effect.

[0004] In recent years, unsupervised learning techniques have shown great potential in fault detection, especially deep learning-based unsupervised methods such as autoencoders and generative adversarial networks (GANs), which can effectively extract the potential representation of data without relying on labeled data. However, although deep learning-based unsupervised methods can effectively extract the potential representation of data in general scenarios, their effectiveness 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 scene: the operating environment of the wind turbine is filled with high-intensity, wide-band 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 impact) are often extremely weak and transient, with their energy and characteristic frequency band easily being overwhelmed or confused by strong background noise. In addition, the rotational speed of the wind turbine blade is continuously changing due to wind speed, resulting in significant non-stationarity and time-varying of its acoustic characteristics (such as pass frequency and harmonics). Existing unsupervised methods often struggle to robustly capture noise and weak fault signals, or fail to adaptively capture and represent these transient fault features that drift with operating conditions and vary in shape, limiting their reliability and accuracy in actual wind turbine blade fault detection.

[0005] Therefore, it is of great theoretical and engineering significance to develop a fault sample-free detection method that can adapt to strong noise and variable operating conditions and is sensitive to early faults. SUMMARY

[0006] The technical problem to be solved by the present application is to provide an unsupervised wind turbine blade fault detection method based on a phase-aware parallel attention mechanism, which solves the problem of strong dependence on labeled data in existing wind turbine blade fault detection methods, insufficient generalization ability under strong noise and variable operating conditions, and difficulty in robustly capturing weak transient fault signals and dynamically changing features.

[0007] The technical solution adopted by the present application to solve the above technical problem is:

[0008] The unsupervised wind turbine blade fault detection method based on the 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, a decoder, an auxiliary encoder, a discriminator, and an anomaly scoring module. The encoder, decoder, and auxiliary encoder each have a phase-aware parallel attention module built-in. 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 phase information in the input features, and acts on the input features to capture time-dimension transient features.

[0009] The low-frequency branch obtains the dimension-reduced features after the input features are dimension-reduced through the average pooling layer, generates attention weights according to the phase information in the input features, and acts on the dimension-reduced features to capture the global features in the frequency dimension;

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

[0011] The detection method comprises the following steps:

[0012] S1. Collecting audio signals of the wind turbine blade during operation;

[0013] S2. Preprocessing the collected audio signals, and extracting double-channel time-frequency features containing amplitude spectrum and phase spectrum by using an improved short-time Fourier transform;

[0014] S3. After the double-channel time-frequency features are preliminarily processed by the convolution unit in the encoder, the phase-aware parallel attention module built-in the encoder is used for processing to generate primary latent representation;

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

[0016] S5. The reconstruction representation is input into the auxiliary encoder, first processed by the convolution unit in the auxiliary encoder, and then processed by the phase-aware parallel attention module built-in the auxiliary encoder to generate secondary latent representation;

[0017] S6. The anomaly score module calculates the anomaly score based on the deviation between the primary latent representation and the secondary latent representation and the deviation between the reconstruction representation and the double-channel time-frequency features;

[0018] S7. According to the extreme value theory, the decision threshold is dynamically set, and if the anomaly score is greater than the threshold, it is determined that the wind turbine blade has a fault.

[0019] Further, the phase-aware parallel attention module comprises a high-frequency branch and a low-frequency branch, and each of the high-frequency branch and the low-frequency branch is provided with a phase-aware multi-head self-attention module, the phase-aware multi-head self-attention module comprises a plurality of parallel attention heads, each of the attention heads comprises three parallel branches: two phase extraction branches with the same structure and a value vector generation branch, and the process of attention calculation of the phase-aware multi-head self-attention module comprises:

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

[0021] The two-dimensional fast Fourier transform is performed on the input feature through two phase extraction branches, and phase information is obtained after convolution processing;

[0022] The value vector is obtained by performing convolution and depth convolution on the input feature through the value vector generation branch;

[0023] The phase information generated by the two phase extraction branches is respectively taken as a query vector and a key vector, and a weight matrix is calculated;

[0024] The weight matrix is applied to the value vector to generate a weighted feature;

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

[0026] Further, in step S2, the collected audio signal is preprocessed, and an improved short-time Fourier transform is used to extract a double-channel time-frequency feature containing an amplitude spectrum and a phase spectrum, which specifically includes:

[0027] The collected audio signal is pre-emphasized and framed according to a 40ms frame length;

[0028] A complex window function is applied to each audio frame after framing , wherein is a Gaussian window function, is a real-time estimated wind turbine blade fundamental frequency; the window length used is dynamically calculated as , wherein is the sampling frequency of the audio signal;

[0029] Fast Fourier transform is performed on the windowed audio frame to obtain a complex spectrum;

[0030] The amplitude spectrum and the phase spectrum are separated from the complex spectrum , and a double-channel feature is output, wherein represents the dimension number of the data, i.e., the channel dimension is 2, the time dimension is T, and the frequency dimension is F.

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

[0032] a. Collect audio signals of wind turbine blades during normal operation to construct a training data set containing only non-fault samples;

[0033] b. Preprocess the audio signals in the training data set, and use an improved short-time Fourier transform to extract a double-channel time-frequency feature containing an amplitude spectrum and a phase spectrum;

[0034] c. obtaining a primary latent representation by using an encoder based on the dual-channel time-frequency feature;

[0035] d. obtaining a reconstruction representation by using a decoder based on the primary latent representation;

[0036] e. obtaining a secondary latent representation by using an auxiliary encoder based on the reconstruction representation;

[0037] f. taking the dual-channel time-frequency feature and the reconstruction representation as inputs of a discriminator, calculating an adversarial loss according to an output of the discriminator;

[0038] g. calculating a composite loss function according to the adversarial loss, a reconstruction error loss, a latent representation consistency loss and a phase consistency loss;

[0039] h. performing a backward gradient update of the model according to the composite loss function to optimize network parameters;

[0040] i. repeating steps b-h until the model parameters converge, obtaining a trained fault detection model.

[0041] Further, in step g, the composite loss function is represented as:

[0042] ;

[0043] wherein, is a total loss of the model; , , are weights of the preset , , ;

[0044] is a reconstruction error loss, ;

[0045] wherein, is a dual-channel time-frequency feature, is a reconstruction representation;

[0046] is a latent representation consistency loss, ;

[0047] wherein, is a primary latent representation, is a secondary latent representation;

[0048] is an adversarial loss, ;

[0049] wherein, represents an implicit function of the discriminator, denotes a Sigmoid function;

[0050] is a phase consistency loss,

[0051] wherein, is a temporal attention weight, is a frequency attention weight, denotes a gradient calculation, denotes a tensor product, denotes a total variation calculation, is a coefficient, taken as 0.5.

[0052] Further, in step S6, the anomaly score module calculates the anomaly score based on the deviation of the primary latent representation and the secondary latent representation and the deviation of the reconstructed representation and the dual-channel time-frequency feature, and the manner of calculating the anomaly score comprises:

[0053]

[0054] wherein, is an anomaly score, is a dynamic weight, is a feature distance score, is a reconstruction error score;

[0055]

[0056]

[0057] wherein, is a phase unwrapping operation; , are the amplitude spectrum and the phase spectrum of the input audio signal, respectively; , are the reconstructed representation and its phase spectrum, respectively; is a preset parameter, taken as 0.3; denotes the inverse matrix of the covariance matrix.

[0058] Further, the dynamic weight is calculated based on the ambient noise level, and specifically comprises:

[0059]

[0060] wherein, is an estimated signal-to-noise ratio, is an adjustment parameter, and the noise spectrum is estimated by a silent segment, is the square of the Frobenius norm of a matrix.

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

[0062] First, the mean value of the abnormal score in the sliding window And the standard deviation ;

[0063] Then, the extreme value coefficient is calculated according to the formula , wherein The preset value is 0.99;

[0064] Finally, the decision threshold Is calculated.

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

[0066] Further, in step S1, the way of collecting the audio signal of the wind turbine blade during operation of the wind turbine blade to be detected includes:

[0067] By deploying a microphone array on the wind turbine blade to be detected, the audio signal of the wind turbine blade during operation is collected, and the collected audio signal is sent to a remote intelligent terminal through wireless transmission, and the intelligent terminal is deployed with a trained fault detection model.

[0068] The beneficial effects of the present application are:

[0069] (1) The fault detection model used in the present application is based on an encoder-decoder-encoder architecture, and a phase-aware parallel attention module is innovatively introduced into the encoder and decoder. The module can accurately capture the weak transient abnormal features generated by the blade fault from the time dimension and identify the feature band drift caused by the speed fluctuation from the frequency dimension under strong background noise (such as wind noise and mechanical noise) through a parallel double-branch architecture (high frequency / low frequency) and a design that explicitly integrates phase information into attention calculation, thereby more robustly modeling the potential fault representation of the wind turbine blade in a complex dynamic operating environment, and further improving the reliability and accuracy in wind turbine blade fault detection.

[0070] (2) The model in the present application adopts an offline unsupervised training method in the training method, which only needs normal working state samples to complete model training, and gets rid of the dependence on scarce fault labeled data. At the same time, in the feature learning stage, the discriminator is introduced based on the generator network for adversarial training, and the phase consistency loss function is combined to constrain the smoothness of the attention weight in the time-frequency dimension, ensuring that the learned features conform to the physical laws of sound wave propagation, thereby optimizing the audio feature space distribution of the wind turbine blade under normal conditions, improving the sensitivity to abnormal features, and enabling the model to maintain stable performance under strong noise and variable working conditions. ​

[0071] (3) The application adopts a multi-scale anomaly scoring algorithm to comprehensively construct a composite detection index by combining feature space distance and reconstruction error when applying the trained model to actual fault detection, dynamically sets a decision threshold combining extreme value theory, realizes high-precision fault early warning, and significantly improves the fault detection rate. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 The figure is a schematic diagram of the fault detection model architecture in the embodiment of the application.

[0073] Figure 2 The figure is a schematic diagram of the application scenario in the embodiment of the application.

[0074] Figure 3 The figure is a schematic diagram of the generative adversarial network structure in the embodiment of the application.

[0075] Figure 4 The figure is a schematic diagram of the phase-aware parallel attention mechanism in the embodiment of the application.

[0076] Figure 5 The figure is a schematic diagram of the phase-aware multi-head self-attention mechanism in the embodiment of the application. DETAILED DESCRIPTION

[0077] The application aims to provide an unsupervised wind power equipment blade fault detection method based on a phase-aware parallel attention mechanism, solve the problem that existing wind power equipment blade fault detection methods have strong dependence on labeled data, have insufficient generalization ability under strong noise and variable working conditions, and are difficult to robustly capture weak transient fault signals and dynamic change characteristics. The core idea is as follows: in view of the characteristics of weak, transient, easily submerged by noise and non-stationary time-varying spectrum of wind turbine blade fault acoustic signals, the sensitivity to subtle waveform distortion is enhanced through a phase-aware mechanism, dynamic STFT is adapted to speed changes, and multi-scale anomaly scoring fusion (feature deviation and signal distortion) and other core designs are adopted, which significantly improves the detection accuracy, robustness and generalization ability of early blade faults (such as micro-cracks and foreign object impact) under unlabeled data conditions, and provides more reliable technical support for intelligent operation and maintenance of wind power equipment.

[0078] The more specific means adopted by the application to realize the above core idea include but are not limited to:

[0079] (1) The built detection model is based on an encoder-decoder-encoder architecture, and a phase-aware parallel attention module is innovatively introduced in the encoder and the decoder. The module can accurately capture weak transient abnormal characteristics (such as crack impact waveforms) generated by blade faults from the time dimension and identify feature 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), so as to more robustly model the potential fault representation of the wind turbine blade in a complex dynamic operating environment.

[0080] (2) In the pre-processing of the collected audio signals of the wind turbine blade in operation, an improved short-time Fourier transform is used. This process uses a complex window function to perform windowing processing on the audio signal, effectively focusing and enhancing the weak fault signal related to the blade rotation, and simultaneously outputting amplitude and phase two kinds of atlas information, providing more rich and targeted time-frequency feature representation for subsequent analysis. In the windowing process, the window length is dynamically adjusted according to the real-time estimated blade rotation frequency, ensuring that the complete periodic characteristics of the blade rotation can be captured each time.

[0081] (3) The model uses an offline unsupervised training method in the training method, which only needs normal state samples to complete model training, and gets rid of the dependence on scarce fault labeled data. At the same time, in the feature learning stage, the introduced discriminator is trained against the generator network, and the phase consistency loss function is used to constrain the smoothness of the attention weight in the time-frequency dimension, ensuring that the learned features conform to the physical laws of sound wave propagation, thereby optimizing the audio feature space distribution of the wind turbine blade under normal state, improving the sensitivity to abnormal features, and enabling the model to maintain stable performance under strong noise and variable operating conditions.

[0082] (4) In the 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 abnormal pattern), while the reconstruction error score directly captures the physical distortion of the original signal (including amplitude error and phase error of waveform distortion containing energy anomaly), thereby revealing the fault from different dimensions. Combined with the extreme value theory, a dynamic decision threshold is set for fault discrimination. The setting of the dynamic decision threshold fully considers the uncertainty in the operation of the wind power equipment, thereby realizing high-precision fault warning and significantly improving the fault detection rate.

[0083] The scheme of the present application will be further described below in combination with the drawings and examples.

[0084] The embodiment provides an unsupervised wind power equipment blade fault detection method based on a phase-aware parallel attention mechanism.

[0085] The fault detection model is a deep adversarial autoencoder architecture, as shown in Figure 1 , which comprises an encoder E1, a decoder D1, an auxiliary encoder E2, a discriminator and an anomaly score module; in the training stage, the encoder E1, the decoder D1, the auxiliary encoder E2 and the discriminator constitute an adversarial network, as shown in Figure 3 ; by performing feature extraction on the training samples (audio data of the wind power equipment blade in the historical normal working state), double-channel time-frequency features containing amplitude spectrum and phase spectrum are obtained, the features are processed by the encoder E1 to obtain primary latent representation, the primary latent representation is input into the decoder D1 to obtain reconstruction representation, and the reconstruction representation is input into the auxiliary encoder E2 to obtain secondary latent representation; based on the double-channel time-frequency features and the reconstruction representation as the input of the discriminator, the discriminator distinguishes the authenticity of the original spectrum and the reconstructed spectrum, thereby calculating the adversarial loss, then combining the reconstruction error loss, the latent representation consistency loss and the phase consistency loss, calculating the composite loss function, and performing reverse gradient update on the model parameters, thereby iteratively optimizing the model parameters.

[0086] In Figure 3 the illustrated structure, the input feature X is a double-channel time-frequency feature obtained by preprocessing, including amplitude spectrum and phase spectrum . The encoder E1 is sequentially connected by a plurality of convolution modules (convolution module 1, convolution module 2, convolution module 3) and a phase-aware parallel attention module; the convolution module gradually compresses the feature dimension through convolution operation, and the phase-aware parallel attention module enhances the time-frequency key information of the convolution feature, and finally outputs the primary latent representation . The decoder D1 adopts a structure symmetrical to the encoder E1, and comprises a phase-aware parallel attention module and a plurality of deconvolution modules (deconvolution module 1, deconvolution module 2, deconvolution module 3); the phase-aware parallel attention module enhances the time-frequency key information of the primary latent representation , and then the deconvolution module gradually restores the dimension to obtain the reconstruction representation . The auxiliary encoder E2 has a structure similar to the encoder E1, which is composed of a plurality of convolution modules (convolution module 1, convolution module 2, convolution module 3) and a phase-aware parallel attention module, which receives the reconstruction representation output by the decoder D1, gradually compresses the feature dimension through convolution operation, and then uses the phase-aware parallel attention module to enhance the time-frequency key information of the convolution feature to generate the secondary latent representation . The discriminator is used to distinguish the input feature with the reconstructed representation , whose input is connected with the original input feature and the reconstructed representation , outputs the discriminative result for calculating the adversarial loss. In addition, there is a skip connection between the corresponding convolutional module and deconvolutional module of the encoder E1 and the decoder D1, which directly transmits low-level feature information to assist in improving the reconstruction accuracy. The entire network realizes accurate learning of the normal state feature distribution through the consistency constraint of the primary latent representation and the secondary latent representation , the error constraint of the input feature and the reconstructed representation , and the adversarial constraint of the discriminator.

[0087] The structure of the phase-aware parallel attention module is shown in Figure 4 , which adopts a parallel double-branch architecture: the high-frequency branch directly inputs the dual-channel time-frequency feature 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 through the same attention module after dimension reduction by average pooling of the dual-channel time-frequency feature, and finally, the output features of the two branches are fused by concatenation and projection to form enhanced representations with high-frequency details and low-frequency structures.

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

[0089] The high-frequency branch directly inputs the input feature into the phase-aware multi-head self-attention module to obtain high-frequency branch attention feature , then performs concatenation and projection (weight matrix ) operations on to generate high-frequency branch output feature ; the low-frequency branch first performs average pooling on the input feature to obtain pooling feature , then also goes through the phase-aware multi-head self-attention to obtain low-frequency branch attention feature , then performs concatenation and projection (weight matrix ) on to generate low-frequency branch output feature . Finally, the output features of the two branches and are concatenated to obtain concatenated feature , and the final projection layer (weight matrix ) is used to fuse to obtain the output feature of the module .

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

[0091] Phase extraction branch 1: after two-dimensional fast Fourier transform is performed on the input, 1x1 convolution (compressing channel dimension), depth convolution (extracting local features) and shape regularization (adjusting matrix dimension) operations are sequentially performed, and phase information Φ used to generate a query vector is output q .

[0092] Phase extraction branch 2: consistent with the structure of the phase extraction branch 1, but has different learnable projection weights from the phase extraction branch 1, after two-dimensional fast Fourier transform is performed on the input, 1x1 convolution (compressing channel dimension), depth convolution (extracting local features) and shape regularization (adjusting matrix dimension) operations are sequentially performed, and phase information Φ used to generate a key vector is output k .

[0093] Value vector generation branch: the input directly undergoes 1x1 convolution, depth convolution and shape regularization operations, and a value vector V is output, which is used to carry feature content.

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

[0095] The module can accurately capture the phase distortion related to the fault in the time-frequency feature by explicitly extracting the phase information and integrating it into the attention mechanism, and can enhance the recognition ability of subtle fault signals.

[0096] In the inference stage, the audio data of the wind turbine blade to be detected is collected, and the feature extraction is performed to obtain the dual-channel time-frequency features containing the amplitude spectrum and the phase spectrum. The features are processed by the encoder E1 to obtain the primary latent representation. The primary latent representation is input into the decoder D1 to obtain the reconstructed representation. The reconstructed representation is input into the auxiliary encoder E2 to obtain the secondary latent representation. The anomaly scoring module calculates the anomaly score based on the deviation between the primary latent representation and the secondary latent representation and the deviation between the reconstructed representation and the dual-channel time-frequency features. The anomaly score is compared with the decision threshold dynamically set according to the extreme value theory. If the anomaly score is greater than the threshold, it is determined that the wind turbine blade has a fault.

[0097] One application scenario of the above fault detection model is as follows Figure 2As shown, audio data during the rotation of the blade is collected by deploying high-precision sound collection devices on the blade of the wind power equipment. In an exemplary embodiment, audio signals generated during the operation of the blade are collected by deploying a high-precision microphone array on the blade of the wind power equipment, the sampling frequency is ≥ 44.1 kHz, the target sound pressure is enhanced by time delay beam forming, and then the collected audio data is transmitted to a remote intelligent terminal through wireless transmission. The intelligent terminal is deployed with a trained fault detection model. When receiving the uploaded audio data, the model is used to process the audio data to detect whether the current wind power equipment blade is faulty.

[0098] In an exemplary embodiment, the fault detection scheme of the wind power equipment blade by the intelligent terminal using the fault detection model includes the following processing process:

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

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

[0101] Specifically, it includes:

[0102] The collected audio signal is pre-emphasized and framed according to a 40ms frame length;

[0103] A composite window function is applied to each audio frame after framing , wherein is a Gaussian window function, is a real-time estimated fundamental frequency of the wind power equipment blade; the window length used is dynamically calculated as , wherein is the sampling frequency of the audio signal;

[0104] Fast Fourier transform is performed on the windowed audio frame to obtain a complex spectrum;

[0105] The amplitude spectrum and the phase spectrum are separated from the complex spectrum, and the double-channel features are outputted, wherein indicates the dimension number of the data, that is, 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 used in this embodiment and the ordinary short-time Fourier transform is its high adaptability to the speed change of the wind turbine blade. The ordinary method usually uses a fixed length time window for analysis, while this method dynamically adjusts the window length according to the real-time estimated blade rotation frequency, ensuring that each analysis can capture the complete period characteristics of the blade rotation; at the same time, it adopts a special composite window design, which effectively focuses and enhances the weak fault signals related to the blade rotation, and simultaneously outputs both amplitude and phase spectrum information, providing more rich and targeted time-frequency feature representation for subsequent analysis. The obtained dual-channel features have phase sensitivity, which can effectively suppress the spectrum leakage problem caused by speed fluctuation, while explicitly preserving the complete physical information of the amplitude spectrum and the phase spectrum, 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 preliminarily processed by the convolution unit in the encoder, and then input into the built-in phase-aware parallel attention module for processing to generate a primary latent representation .

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

[0110] In this step, the primary latent representation is input into the decoder, first processed by the built-in phase-aware parallel attention module, and then processed by the deconvolution unit in the decoder, to reconstruct a reconstructed representation with the same dimension as the dual-channel time-frequency features .

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

[0112] In this step, the reconstructed representation is input into the auxiliary encoder, first processed by the convolution unit, and then processed by the built-in phase-aware parallel attention module to generate a secondary latent representation .

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

[0114] In this step, 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 .

[0115] In this embodiment, when calculating the anomaly score, a multi-scale anomaly scoring algorithm is used, which mainly considers the feature distance score and the reconstruction error score. Among them, the feature distance score calculates the primary potential representation and secondary latent representations The standardized Mahalanobis distance reflects the state offset of the latent space (systematic abnormal patterns). The reconstruction error score combines the amplitude error and the unwrapping phase error to directly capture the physical distortion of the original signal (including the amplitude error of energy anomalies and the phase error of waveform distortion). Finally, the two scores are fused through dynamic weighting to fully reveal the fault from different dimensions.

[0116] Feature distance score The calculation method is:

[0117] ;

[0118] in, It means to find the inverse matrix of the covariance matrix.

[0119] Reconstruction error score The calculation method is:

[0120] ;

[0121] in, It is a phase unwrapping operation; are the amplitude and phase spectra of the input audio; and are the reconstructed representation and its phase spectrum, , is a preset parameter.

[0122] Final rating The calculation method is:

[0123] ;

[0124] in, It is a dynamic weight that is adaptively adjusted according to the environmental signal-to-noise ratio. The specific calculation method includes:

[0125] ;

[0126] in, is the estimated signal-to-noise ratio; is the adjustment parameter; noise spectrum Estimation through silent segments; is the squared Frobenius norm of the matrix, which is the sum of the squares of the absolute values ​​of all the elements of the matrix.

[0127] In this embodiment, through the environment-adaptive dynamic weight fusion, the reliability of the two can be balanced under different signal-to-noise ratios, that is, the reconstructed details are emphasized when the signal-to-noise ratio is high, and the feature distribution is relied on when the signal-to-noise ratio is low.

[0128] S6. Determine whether the wind power equipment blade has a fault based on the dynamic decision threshold:

[0129] In this step, the abnormal score calculated in step S5 is compared with the decision threshold set dynamically according to the extreme value theory. If the abnormal score is greater than the threshold, it is determined that the wind power equipment blade has a fault.

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

[0131] First, the mean value and the standard deviation of the composite abnormal score in the sliding window (the data time range used for dynamically calculating the abnormal score in the real-time monitoring process of the system, usually set as an integer multiple of the complete rotation period of the wind power blade) are obtained.

[0132] Then, the extreme value coefficient is calculated according to the formula (wherein takes a preset value of 0.99).

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

[0134] When the value of the real-time calculated abnormal score exceeds , a fault alarm is triggered.

[0135] By dynamically setting the decision threshold by the extreme value theory, the robustness and accuracy of the detection can be significantly improved.

[0136] The training process of the fault detection model is introduced as follows:

[0137] The fault detection model in this embodiment is trained in an offline unsupervised manner, and the training process includes:

[0138] a. Collect the audio signals of the wind power equipment blade during normal operation, and construct a training data set containing only non-fault samples;

[0139] b. Preprocess the audio signals in the training data set, and extract the dual-channel time-frequency features containing the amplitude spectrum and the phase spectrum by using the improved short-time Fourier transform;

[0140] c. Based on the dual-channel time-frequency features, the primary latent representation is obtained by using the encoder. ​​​​​​

[0141] d. obtaining a reconstructed representation using a decoder based on the primary latent representation;

[0142] e. obtaining a secondary latent representation using an auxiliary encoder based on the reconstructed representation;

[0143] f. taking the dual-channel time-frequency feature and the reconstructed representation as inputs of a discriminator, and calculating an adversarial loss according to an output of the discriminator;

[0144] g. calculating a composite loss function according to the adversarial loss, a reconstruction error loss, a latent representation consistency loss, and a phase consistency loss;

[0145] h. performing a backward gradient update of the model according to the composite loss function to optimize network parameters;

[0146] i. repeating steps b-h until the model parameters converge, and obtaining a trained fault detection model.

[0147] Based on the above training method, only normal state samples are needed to complete model training, and the dependence on scarce fault labeled data is eliminated. Meanwhile, in the feature learning stage, the discriminator introduced is trained in an adversarial manner with the generator network, and the phase consistency loss function is used to constrain the smoothness of the attention weight in the time-frequency dimension, so as to ensure that the learned features conform to the physical laws of sound wave propagation, thereby optimizing the audio feature space distribution of the wind turbine blade under normal conditions, improving the sensitivity to abnormal features, and enabling the model to maintain stable performance under strong noise and variable working conditions.

[0148] Specifically, the embodiment adopts a joint optimization of four types of loss functions in the model training stage: a reconstruction error loss , a latent representation consistency loss , an adversarial loss , and a phase consistency loss .

[0149] Among them, the reconstruction error loss ensures that the model can accurately learn the details of the time-frequency feature under normal working conditions (such as amplitude distribution rules and phase continuity) by directly constraining the reconstructed representation generated by the decoder to be consistent with the original input feature (including amplitude spectrum and phase spectrum). This constraint forces the decoder to faithfully restore the key information of the input feature and reduces the reconstruction deviation under normal conditions. The calculation is as follows:

[0150] ;

[0151] Among them, is the dual-channel time-frequency feature, is the reconstructed representation.

[0152] Potential representation consistency loss By constraining the primary potential representation generated by the encoder to be consistent with the secondary potential representation generated by the auxiliary encoder in distribution, it is ensured that the features reconstructed by the decoder can still accurately map back to the original potential feature space after passing through the auxiliary encoder. This constraint strengthens the model's stable representation ability for normal working condition features, reduces the loss or distortion of feature information in the reconstruction process, and makes the model more accurately learn the feature distribution regularity of the normal state, so as to effectively identify the abnormal state by the deviation between the two in the detection stage. The calculation is as follows:

[0153] ;

[0154] wherein, is the primary potential representation, is the secondary potential representation.

[0155] Adversarial loss By discriminating the true and false of the original dual-channel time-frequency features and the reconstructed features by the discriminator, the generator generates reconstructed features closer to the real physical characteristics. This loss strengthens the authenticity of the reconstructed features in the time-frequency domain (such as the rationality of the amplitude spectrum distribution and the physical continuity of the phase spectrum), avoiding the model simply copying the features to complete the reconstruction, but learning the essence of the features in the normal working condition, thereby improving the sensitivity to the abnormal features caused by subtle faults. The calculation is as follows:

[0156] ;

[0157] wherein, represents the implicit function of the discriminator, represents the Sigmoid function.

[0158] Phase consistency loss By combining the physical acoustic characteristics with the deep learning attention mechanism through mathematical formulas, the consistency of the original dual-channel time-frequency features and the reconstructed features in the phase spectrum is constrained, including the numerical deviation after phase spectrum unwrapping and the derivative continuity deviation of the phase with respect to time / frequency. The core role is to ensure that the phase information of the reconstructed features conforms to the physical law (such as the phase change trend corresponding to the blade rotation fundamental frequency), avoiding the modeling deviation of the normal state caused by phase distortion. The accurate restoration of phase information can enhance the model's ability to capture phase mutations caused by faults (such as the vibration phase shift caused by blade cracks), supplementing the deficiency of amplitude spectrum information. The calculation is as follows:

[0159] ;

[0160] wherein, is the time attention weight, is the frequency attention weight, represents gradient calculation, denotes the tensor product, denotes the total variation calculation, is the coefficient, taking 0.5.

[0161] In the above formula, the total variation regularization term is the smoothness constraint of the time-frequency attention. This term calculates the gradient and the frequency attention weight respectively, and then constructs the time-frequency joint gradient matrix through the tensor product , and then calculates the total variation . This makes the attention weight change smoothly in the time and frequency dimensions, avoiding sharp fluctuations due to noise interference. Finally, the composite loss function is represented as:

[0162]

[0163] ;

[0164] wherein, is the total loss of the model; , , are the preset weights of , , .

[0165] In an exemplary embodiment, the training process adopted is as follows:

[0166] First, a special training data set is constructed: continuous audio during normal operation of wind power equipment is collected, and the original waveform with a duration of not less than 200 hours is obtained through a microphone array.

[0167] The data preprocessing adopts the process described in the foregoing: the original signal is pre-emphasized to enhance the high-frequency components, after being framed according to a frame length of 40 ms (50% overlap rate), a dynamic window length STFT transformation is applied using a complex window function to generate a dual-channel time-frequency pattern sample set containing amplitude spectrum and phase spectrum, and finally about 150,000 training samples are formed.

[0168] The key training control measures include: verifying the reconstruction signal-to-noise ratio after each training, and triggering gradient clipping when the signal-to-noise ratio is lower than 35 dB; the phase consistency loss weight linearly increases with the training round, and the maximum proportion reaches 40% of the total loss. The Adam optimizer is used for back propagation, the initial learning rate is set to 5×10⁻ 4 , and is attenuated by 30% every 50 rounds, a total of 300 rounds.

[0169] The entire training process takes about 36 hours to complete on one NVIDIA A5000 GPU, only consuming normal operating data, without any fault annotation.​

[0170] Verification example:

[0171] To verify the accuracy of the wind turbine blade fault detection method of the present application, the detection accuracy of three different wind turbine blade detection models will be compared in this verification example. The audio data tested 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 by the present application, the CNN (Convolutional Neural Network) model and the CRNN (Convolutional Recurrent Neural Network) model in the prior art.

[0172] Among them, the detection accuracy of the model proposed by the present application, the detection accuracy of the CNN model and the detection accuracy of the CRNN model are shown in Table 1, Table 2 and Table 3 respectively.

[0173] Table 1 Detection accuracy table of the model proposed by the present application

[0174] Table 2 Detection accuracy table of the CNN model

[0175] Table 3 Detection accuracy table of the CRNN model

[0176] In the above table, f1, f2 and f3 represent three different wind turbines used in land, sea and high altitude areas respectively. Part 1 is the blade root bolt connection area, part 2 is the front edge wind erosion protection layer, and part 3 is the rear edge main beam cap area.

[0177] From the data, the average detection accuracy of the method proposed by the present application on devices f1, f2 and f3 is 85.54%, 87.43% and 84.35% respectively, which is significantly higher than the average detection accuracy of the CNN and CRNN models on these three devices. As for each part, the accuracy of the present application method in the same part is also higher than that of the other two models. This shows that the method proposed by the present application has better accuracy and reliability in wind turbine blade fault detection, and can more effectively identify the fault condition of the blade, providing more powerful support for the maintenance and management of wind turbines.

[0178] Therefore, although the above has described the embodiments of the present application, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and all of them do not depart from the protection scope of the present application.

Claims

1. An unsupervised wind turbine blade fault detection method based on a phase-aware parallel attention mechanism, which adopts 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, and the encoder, decoder and auxiliary encoder each have a phase-aware parallel attention module built-in; 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 according to phase information in the input features, and acts on the input features to capture time-dimension transient features; the low-frequency branch obtains reduced features after reducing the input features through an average pooling layer, generates attention weights according to phase information in the input features, and acts on the reduced features to capture frequency-dimension global features; the time-dimension transient features output by the high-frequency branch and the frequency-dimension global features output by the low-frequency branch are fused to obtain enhanced features output by the phase-aware parallel attention module; the detection method includes the following steps: S1. Collect the audio signal of the wind turbine blade during operation; S2. Preprocess the collected audio signal, and extract dual-channel time-frequency features containing amplitude spectrum and phase spectrum using an improved short-time Fourier transform; S3. After the dual-channel time-frequency features are preliminarily processed by the convolution unit in the encoder, they are input into the phase-aware parallel attention module built-in to generate primary latent representations; S4. The primary latent representations are input into the decoder, first processed by the phase-aware parallel attention module built-in, and then processed by the deconvolution unit in the decoder to reconstruct representations consistent with the dimensions of the dual-channel time-frequency features; S5. The reconstructed representations are input into the auxiliary encoder, first processed by the convolution unit, and then processed by the phase-aware parallel attention module built-in to generate secondary latent representations; S6. The anomaly scoring module calculates the anomaly score based on the deviation between the primary latent representations and the secondary latent representations, and the deviation between the reconstructed representations and the dual-channel time-frequency features; S7. According to the extreme value theory, the decision threshold is dynamically set, and if the anomaly score is greater than the threshold, it is determined that the wind turbine blade has a fault.

2. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism according to claim 1, wherein, In the phase-aware parallel attention module, there are phase-aware multi-head self-attention modules in the high-frequency branch and the low-frequency branch, and the phase-aware multi-head self-attention module includes multiple parallel attention heads, each of which includes three parallel branches: two phase extraction branches with the same structure and a value vector generation branch. The process of attention calculation performed by the phase-aware multi-head self-attention module includes: (1) Each attention head calculates the weighted features by the following method: Perform two-dimensional fast Fourier transform on the input features through the two phase extraction branches, and obtain phase information after convolution processing; obtain a value vector by performing convolution and depth convolution on the input features through the value vector generation branch; use the phase information generated by the two phase extraction branches as the query vector and the key vector respectively to calculate the weight matrix; The weight matrix is applied to the value vector to generate a weighted feature; (2) The weighted features output by all attention heads are spliced and projected to obtain a final feature representation 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 according to claim 1, wherein, In step S2, the collected audio signal is preprocessed, and an improved short-time Fourier transform is used to extract a dual-channel time-frequency feature containing an amplitude spectrum and a phase spectrum. Specifically, the steps include: The collected audio signal is pre-emphasized and framed according to a 40ms frame length; applying a complex window function to each audio frame after framing wherein, is a Gaussian window function, is a real-time estimated fundamental frequency of the wind turbine blade; the employed window length is dynamically calculated as wherein, is the audio signal sampling frequency; The windowed audio frame is subjected to a fast Fourier transform to obtain a complex spectrum; Separating the magnitude spectrum from the complex spectrum and the phase spectrum and outputting a two-channel feature wherein, denotes the number of dimensions of the data, i.e. 2 for the channel dimension, T for the time dimension and F for the frequency dimension.

4. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism of claim 1, wherein, The fault detection model is trained in an offline unsupervised manner, and the training process includes: a. Collect audio signals of wind turbine blades during normal operation to construct a training data set containing only non-fault samples; b. Preprocess the audio signals in the training data set, and use an improved short-time Fourier transform to extract a dual-channel time-frequency feature containing an amplitude spectrum and a phase spectrum; c. Based on the dual-channel time-frequency feature, use an encoder to obtain a primary latent representation; d. Based on the primary latent representation, use a decoder to obtain a reconstructed representation; e. Based on the reconstructed representation, use an auxiliary encoder to obtain a secondary latent representation; f. Take the dual-channel time-frequency feature and the reconstructed representation as inputs of a discriminator, and calculate an adversarial loss according to the output of the discriminator; g. Calculate a composite loss function based on the adversarial loss, the reconstruction error loss, the latent representation consistency loss, and the phase consistency loss; h. Perform backward gradient update of the model according to the composite loss function to optimize the network parameters; i. Repeat steps b-h 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 of claim 4, wherein, In step g, the composite loss function is represented as: ； wherein, is the total loss for the model; , , are preset , , weights, respectively. to reconstruct the error loss, ; wherein, is a two-channel time-frequency feature, is a reconstruction representation; to potentially represent a consistency loss, ; wherein, is a primary latent representation, is a secondary latent representation; To counteract the loss, ; wherein, represents an implicit function of the discriminator, represents a Sigmoid function; for phase consistency loss, ; wherein, is a temporal attention weight, is a frequency attention weight, denotes a gradient computation, denotes a tensor product, denotes a total variation computation, is a coefficient, taken as 0.

5.

6. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism of claim 5, wherein, In step S6, the anomaly score module calculates the anomaly score based on the deviation between the primary latent representation and the secondary latent representation, and the deviation between the reconstructed representation and the dual-channel time-frequency feature. The calculation method includes: ； wherein, is an anomaly score, is a dynamic weight, is a feature distance score, is a reconstruction error score; ； ； wherein, is a phase unwrapping operation; , are the amplitude spectrum and the phase spectrum of the input audio signal, respectively; , are the reconstructed representation and its phase spectrum, respectively; is a preset parameter, taken as 0.3; denotes the inverse of the covariance matrix.

7. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism of claim 6, wherein, Dynamic weights The calculation is based on the ambient noise level, in particular comprising: ； wherein is the estimated signal-to-noise ratio, is the adjustment parameter, the noise spectrum is estimated by the silent segment, is the square of the Frobenius norm of the matrix.

8. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism of claim 7, wherein, In step S7, the decision threshold is dynamically set according to the extreme value theory, including: First, the mean and standard deviation of the anomaly scores within the sliding window are obtained ; Then the extreme value coefficient is calculated according to the formula Wherein, Take the preset value 0.99;​ Finally, the decision threshold is calculated .

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

10. The unsupervised wind turbine blade fault detection method based on phase-aware parallel attention mechanism according to any one of claims 1-9, wherein, In step S1, the audio signal of the wind turbine blade during operation is collected in the following manner: By deploying a microphone array on the wind turbine blade to be detected, the audio signal of the wind turbine blade during operation is collected, and the collected audio signal is sent to a remote intelligent terminal through wireless transmission. The intelligent terminal is equipped with a trained fault detection model.

Images