Wind power blade fault detection method and system based on audio potential characterization

Abstract

The invention relates to a wind turbine generator fault monitoring technology, and discloses a wind turbine blade fault detection method and system based on audio potential characterization, which improve the accuracy of wind turbine blade fault detection and do not depend on annotated data. According to the method, a trained fault detection model is adopted to carry out fault detection on the wind power blade, the fault detection model comprises a generator network adopting an encoder-decoder-encoder architecture, and adversarial training is carried out by introducing a discriminator and the generator network. The detection method comprises the following steps: firstly, collecting an original audio when a blade works, carrying out preprocessing and feature extraction to obtain an original Mel spectrum, then carrying out processing through a first encoder to obtain a first potential representation, carrying out processing through a decoder to obtain a reconstructed Mel spectrum, and then carrying out processing through a second encoder to obtain a second potential representation; and finally, calculating an abnormal condition score according to the two potential representations, and judging whether the wind power blade is abnormal or not according to the score. The method is suitable for wind turbine generator blade fault detection.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbine fault monitoring, and particularly to a method and system for detecting wind turbine blade faults based on audio latent representations. Background Art

[0002] The detection of wind turbine blade faults is a key link in wind power operation and maintenance, directly affecting the operation efficiency and safety of wind farms. In the prior art, blade fault detection technologies can be roughly divided into two categories: one is the detection method based on physical models, which predicts faults by establishing physical models of blades. However, due to the complexity and uncertainty of the operating environment of wind power equipment, physical models are often difficult to accurately reflect the actual operating state, resulting in insufficient detection accuracy. The other is the detection method based on data-driven, especially using machine learning technology for fault detection. However, most of the existing data-driven methods rely on supervised learning, requiring a large amount of labeled data, which is complex to operate and costly.

[0003] For example, in the patent application with the patent publication number CN115406630A and the name "A method for detecting wind turbine blade faults using passive acoustic signals based on machine learning", the following solution is disclosed: in response to a trigger event for detecting faults in a wind turbine blade, an audio signal generated when the wind turbine blade rotates is collected; the Mel-frequency cepstral coefficients of the audio signal are extracted; a fault detection model is called to analyze and process the Mel-frequency cepstral coefficients to obtain a detection result for detecting faults in the wind turbine blade.

[0004] In the above solution, a supervised learning method is adopted, which requires a large amount of labeled training data, and has the problems of complex operation and high cost; moreover, in the field of wind turbine blade fault detection, due to factors such as the complexity of the operating environment of wind turbine blades, multiple fault types, and complex detection, it is very difficult to obtain a sufficient number of high-quality labeled data in practical applications, which limits the training and optimization of the model, thus affecting the detection ability.

[0005] In recent years, unsupervised learning technologies have shown great potential in the field of fault detection. Especially unsupervised methods based on deep learning, such as autoencoders, generative adversarial networks (GANs), etc., can effectively extract the latent representations of data without relying on labeled data. However, the application of existing unsupervised methods in wind turbine blade fault detection still faces the following challenges: (1) Difficulty in processing complex audio signals The audio signals generated during the operation of wind power equipment usually contain a large amount of noise and interference, and the blade fault signals are often relatively weak. Existing unsupervised methods are difficult to effectively extract fault features when processing such complex audio signals, and are easily interfered by noise, resulting in a decrease in detection accuracy.

[0006] (2) Insufficient ability to capture fault characteristics The audio characteristics of wind turbine blade faults may be diverse and complex, and different types of faults may produce different audio patterns. Existing unsupervised methods are insufficient in capturing these subtle and complex fault characteristics.

[0007] (3) Limited model generalization ability The operating environment of wind power equipment is complex and changeable, and the audio signals of different equipment and different locations may vary significantly. Existing unsupervised methods often have difficulty adapting to this diversity and complexity during the training process, resulting in limited model generalization ability and difficulty in maintaining stable detection performance under different operating conditions. Summary of the Invention

[0008] The technical problem to be solved by the present invention is: to propose a method and system for detecting wind turbine blade faults based on audio latent representation, improving the accuracy of wind turbine blade fault detection without relying on labeled data.

[0009] The technical solution adopted by the present invention to solve the above technical problems is as follows: On the one hand, the present invention provides a method for detecting wind turbine blade faults based on audio latent representation, using a trained fault detection model to detect faults in wind turbine blades. The fault detection model includes a generator network adopting an encoder-decoder-encoder architecture, and the generator network includes a first encoder, a decoder, and a second encoder; the detection method includes the following steps: S1. Collect the original audio when the wind turbine equipment blade is working; S2. Preprocess the original audio and extract Mel spectrogram features to obtain the original Mel spectrogram; S3. Based on the original Mel spectrogram, obtain the first latent audio representation using the first encoder; S4. Based on the first latent audio representation, obtain the reconstructed Mel spectrogram using the decoder; S5. Based on the reconstructed Mel spectrogram, obtain the second latent audio representation using the second encoder; S6. Calculate the anomaly score according to the first latent audio representation and the second latent audio representation. If the anomaly score is greater than a preset threshold, it is determined that the wind turbine equipment blade is abnormal.

[0010] Further, in step S1, the collection of the original audio when the wind turbine equipment blade is working includes: Set audio acquisition devices at different positions of the wind turbine equipment, and collect audio data when the wind turbine equipment blade is working through the audio acquisition devices.

[0011] Further, in step S2, the preprocessing of the original audio and the extraction of Mel spectrogram features to obtain the original Mel spectrogram include: Perform pre-emphasis processing on the original audio signal; Segment the pre-emphasized audio signal into frame signals with a length of 20 - 40 milliseconds; Apply a window function to each frame signal for processing; Perform a fast Fourier transform on each frame signal processed by the window function to obtain a short-time spectrum; Use a Mel filter bank to filter the short-time spectrum, convert the linear frequency to Mel frequency, and obtain the original Mel spectrogram.

[0012] Further, both the first encoder and the second encoder include three cascaded convolutional modules, an attention module, and a multi-layer perceptron; the convolutional module is used to perform a convolutional operation on the input Mel spectrogram and perform two-dimensional instance normalization processing to obtain a normalized feature tensor; the attention module is used to calculate attention weights for the normalized feature tensor to obtain a weighted feature representation; the multi-layer perceptron is used to flatten the weighted feature representation and perform a non-linear transformation through a fully connected layer to output an audio characterization vector.

[0013] Further, the decoder includes a multi-layer perceptron, which is used to restore the first latent audio characterization generated by the first encoder to a high-dimensional Mel spectrogram feature to obtain a reconstructed Mel spectrogram.

[0014] Further, in step S6, the calculation of the anomaly score according to the first latent audio characterization and the second latent audio characterization includes: ; where is the anomaly score of the x-th audio, represents the first latent audio characterization of the x-th audio, represents the second latent audio characterization of the x-th audio, and represents the L2 norm.

[0015] Further, the fault detection model is trained in an offline manner and an discriminator is introduced to form an adversarial network. The training process includes: a. Collect the original audio when the wind turbine blade is working; b. Preprocess the original audio and extract Mel spectrogram features to obtain the original Mel spectrogram; c. Based on the original Mel spectrogram, use the generator network in the fault detection model to obtain a reconstructed Mel spectrogram; d. Add random noise to the original Mel spectrogram to obtain a generated original Mel spectrogram; e. Based on the generated original Mel spectrogram, obtain the generated reconstructed Mel spectrogram using the generator network in the fault detection model; f. Use the original Mel spectrogram, the reconstructed Mel spectrogram, the generated original Mel spectrogram, and the generated reconstructed Mel spectrogram as the input of the discriminator, and conduct adversarial training with the generator network; g. Repeat steps a - f until the training termination condition is reached to obtain the trained fault detection model.

[0016] Further, in step f, when conducting adversarial training, the loss function adopted by the fault detection model includes: ; where is the total loss; , , are the first - part loss, the second - part loss, and the third - part loss respectively; are the weights corresponding to the first - part loss, the second - part loss, and the third - part loss respectively; ; ； ; where represents the original Mel spectrogram or the generated original Mel spectrogram; represents the reconstructed Mel spectrogram or the generated reconstructed Mel spectrogram; represents the first latent audio representation obtained after the original Mel spectrogram or the generated original Mel spectrogram is processed by the first encoder; represents the second latent audio representation obtained after the reconstructed Mel spectrogram or the generated reconstructed Mel spectrogram is processed by the second encoder; represents the implicit function of the discriminator; represents the Sigmoid function; represents the L1 norm; represents the L2 norm.

[0017] Further, in step f, when conducting adversarial training, the Adam algorithm is used to update the parameters of the fault detection model, and the update process includes: ; ; ; ; ; ; Among them, represents the time-step gradient at time t; represents gradient calculation; represents the loss when the network parameters are ; is the exponential moving average; is the first-order decay rate of moment estimation; is the corrected gradient mean at time t; represents the second moment at time step t - 1; is the second-order decay rate of moment estimation; represents the corrected gradient square; represents the bias-corrected second-moment gradient estimation; is a constant; is the learning rate.

[0018] On the other hand, the present invention also provides a wind turbine blade fault detection system based on audio latent representation, including a sound collection device, a wireless transmission module, and an intelligent terminal; The sound collection device is used to collect the original audio when the wind turbine blade is working; The wireless transmission module is used to transmit the collected original audio to the intelligent terminal; The intelligent terminal is used to perform fault detection on the wind turbine blade according to the above-mentioned wind turbine blade fault detection method based on audio latent representation.

[0019] The beneficial effects of the present invention are as follows: The present invention adopts an unsupervised learning method, only uses the audio data of the wind turbine blade in the normal operation state for training, does not rely on labeled data, and by introducing adversarial training, enables the model to automatically and effectively learn the distribution characteristics of normal sample data, so that when detecting faults, it can more accurately identify abnormalities. Based on the reconstruction ability of the model, the diversification of data samples can be realized during the training process, thereby improving the generalization ability of the model.

[0020] In addition, the model adopts a new encoder-decoder-encoder architecture. The original Mel spectrogram features are encoded into low-dimensional latent representations by the first encoder, the low-dimensional latent representations are restored to high-dimensional Mel spectrogram features by the decoder, and then the features are further extracted by the second encoder, which can generate richer feature expressions. Adopting this model structure can extract the features of the audio signal from multiple levels, thereby more accurately capturing the fault features and improving the accuracy of fault detection.

[0021] In addition, the attention mechanism is introduced into the model structure, enabling the model to automatically focus on the features related to fault detection, thereby improving the detection accuracy. Description of the Drawings

[0022] Figure 1 This is the flowchart of the wind turbine blade fault detection method based on audio latent representation in the present invention; Figure 2 This is the schematic diagram of the wind turbine blade fault detection system based on audio latent representation in the present invention; Figure 3 This is the comparison chart of the detection accuracy rates when the present invention's solution and the existing model are respectively adopted in the embodiment. Detailed implementation manners

[0023] The present invention aims to provide a wind turbine blade fault detection method and system based on audio latent representation, improve the accuracy of wind turbine blade fault detection, and do not rely on labeled data. Its core idea is: unsupervised learning and adversarial training are introduced, and a new model architecture of encoder-decoder-encoder is adopted. Specifically, the present invention adopts an unsupervised learning method and only uses the audio data in the normal operation state of the wind power equipment for training without any labeled data, which greatly reduces the cost and time of data preparation. Based on the unsupervised learning method, the internal feature distribution of the data can be automatically learned, and it has strong adaptability to the distribution changes of the data. In the actual operation environment of the wind power equipment, due to factors such as environmental noise and equipment aging, the distribution of the audio data may change, and the unsupervised learning method can better adapt to these changes. The model adopts the structure of encoder-decoder-encoder, which can capture the complex features and latent patterns in the audio signal more deeply from multiple levels and improve the richness of audio feature expression. An attention module is introduced in the encoder part, which can capture the global information of the audio signal and the long-distance dependencies within the features from both the time and frequency dimensions, enabling the model to more accurately focus on the features related to faults, thereby improving the detection accuracy. During model training, adversarial training is carried out through the generator network and discriminator of the model. Through adversarial training, the generator network and the discriminator compete with each other. The generator network continuously generates samples closer to the real data, and the discriminator continuously improves its discrimination ability. By continuously optimizing its own parameters, finally the generator network can more accurately and deeply understand the distribution characteristics of the audio data and improve the accuracy of wind turbine blade fault detection.

[0024] In specific implementation, the present invention first deploys a wind turbine blade fault detection system, and its structure is shown in Figure 2, including a sound collection device, a wireless transmission module, and an intelligent terminal; among them, the sound collection device and the wireless transmission module are arranged on the blade of the wind power equipment. To improve the accuracy of detection, multiple high-precision sound collection devices can be installed at different parts of the blade of the wind power equipment. These devices can capture the subtle sound changes generated during the operation of the blade. The audio data collected by the sound collection device is transmitted to the intelligent terminal through the wireless transmission module. The intelligent terminal is arranged remotely and is used to process and analyze the audio data to detect whether there is a fault in the blade. The intelligent terminal can be a desktop computer, a laptop computer, etc.

[0025] The analysis of the audio data by the intelligent terminal depends on the fault detection model deployed on the terminal. The fault detection model includes a generator network adopting an encoder-decoder-encoder architecture, which has been offline trained before being deployed to the terminal. During the offline training process of the model, in order to improve the accuracy and generalization ability of the model for fault detection, a discriminator is introduced to perform adversarial training with the generator network in the model. The process is as follows: a. Collect the original audio when the blade of the wind power equipment is working; In this step, the high-precision sound collection device installed on the blade of the wind power equipment is used to collect the audio data when the blade of the wind power equipment is working. These audio data can contain information such as the original audio, the collection location, the model of the wind power equipment, the sampling rate, and the duration.

[0026] b. Preprocess the original audio and extract the Mel spectrogram features to obtain the original Mel spectrogram; In this step, first, pre-emphasis processing is performed on the original audio signal. The high-frequency components are enhanced through a high-pass filter to balance the spectrum. Then, the pre-emphasized audio signal is segmented into frame signals with a length of 20 - 40 milliseconds. After frame segmentation, each frame signal is processed by applying a window function (such as a Hamming window) to reduce the spectrum leakage at the frame edge. Then, a fast Fourier transform (FFT) is performed on each frame signal to convert the time-domain signal into a frequency-domain signal to obtain a short-time spectrum. Finally, the spectrum is filtered by a Mel filter bank to convert the linear frequency into Mel frequency to obtain the original Mel spectrogram.

[0027] Among them, the relationship between the Mel frequency and the linear frequency is: ; In the formula, f is the linear frequency, is the converted Mel frequency.

[0028] c. Based on the original Mel spectrogram, use the generator network in the fault detection model to obtain the reconstructed Mel spectrogram; In this step, the first encoder in the generator network of the fault detection model extracts features from the original Mel spectrogram to obtain a low-dimensional latent representation, that is, the first latent audio representation; the decoder restores the low-dimensional latent representation to a high-dimensional Mel spectrogram feature to obtain the reconstructed Mel spectrogram; the second encoder further extracts features from the reconstructed Mel spectrogram to obtain a richer latent representation, that is, the second latent audio representation.

[0029] d. Add random noise to the original Mel spectrogram to obtain the generated original Mel spectrogram; In this step, by adding random noise to the original Mel spectrogram, the Mel spectrogram after adding noise, that is, the generated original Mel spectrogram, is obtained as fake sample data.

[0030] e. Based on the generated original Mel spectrogram, use the generator network in the fault detection model to obtain the generated reconstructed Mel spectrogram; In this step, taking the generated original Mel spectrogram as the input of the generator network and using the generator network to process it, the generated reconstructed Mel spectrogram can be obtained, which is also used as fake sample data.

[0031] f. Use the original Mel spectrogram, the reconstructed Mel spectrogram, the generated original Mel spectrogram, and the generated reconstructed Mel spectrogram as the input of the discriminator and perform adversarial training with the generator network; In this step, the data input to the discriminator includes real sample data and fake sample data. Among them, the real sample data includes the original Mel spectrogram and the reconstructed Mel spectrogram, and the fake sample data includes the generated original Mel spectrogram and the generated reconstructed Mel spectrogram.

[0032] The discriminator is composed of several convolutional layers, activation functions, and fully connected layers, and is responsible for distinguishing whether the input sample data is real sample data or fake sample data. During the adversarial training process, the generator and the discriminator compete and promote each other. The generator tries to generate realistic fake data samples to deceive the discriminator, and the discriminator tries to improve the discrimination accuracy and distinguish real samples and fake samples as much as possible. Eventually, the generator network can deeply understand the distribution characteristics of audio data, so that it can accurately reconstruct the audio Mel spectrogram that conforms to the audio data distribution in subsequent actual applications.

[0033] When performing adversarial training, the loss function adopted by the fault detection model includes: ; Among them, is the total loss; , , are the first part of the loss, the second part of the loss, and the third part of the loss respectively; are the weights corresponding to the first part of the loss, the second part of the loss, and the third part of the loss respectively; ; ； ; Among them, represents the original Mel spectrogram or the generated original Mel spectrogram; represents the reconstructed Mel spectrogram or the generated reconstructed Mel spectrogram; represents the first latent audio representation obtained after the original Mel spectrogram or the generated original Mel spectrogram is processed by the first encoder; represents the second latent audio representation obtained after the reconstructed Mel spectrogram or the generated reconstructed Mel spectrogram is processed by the second encoder; represents the implicit function of the discriminator; represents the Sigmoid function; represents the L1 norm; represents the L2 norm.

[0034] In terms of parameter update method, the Adam algorithm is used to update the parameters of the fault detection model, and the update process includes: ; ; ; ; ; ; Among them, represents the time-step gradient at time t; represents the calculation of the gradient; represents the loss when the network parameter is ; is the exponential moving average; is the first decay rate of the moment estimation; is the corrected gradient mean at time t; represents the second moment at time step t - 1; is the second decay rate of the moment estimation; represents the corrected gradient square; represents the bias-corrected second-moment gradient estimation; is a constant; is the learning rate.

[0035] g. Repeat steps a - f until the training termination condition is reached to obtain the trained fault detection model.

[0036] In this step, the training termination condition can be set according to actual needs, such as iterating to a certain number of rounds, or the trained fault detection model converges, etc. The trained fault detection model can be deployed in an intelligent terminal to detect whether the blades of the wind power equipment to be detected are faulty.

[0037] When the above system is actually applied, the flow of the wind turbine blade fault detection method based on audio latent representation provided by the present invention is shown in Figure 1 , and it includes the following steps: S1. Collect the original audio of the wind turbine blade during operation; In this step, the acquisition method of the original audio is as described in the previous step a.

[0038] S2. Preprocess the original audio and extract Mel spectrogram features to obtain the original Mel spectrogram; In this step, the preprocessing and feature extraction process of the original audio are as described in the previous step b.

[0039] S3. Based on the original Mel spectrogram, use the first encoder to obtain the first latent audio representation; In this step, the first encoder includes three cascaded convolutional modules, an attention module, and a multi-layer perceptron; after the original Mel spectrogram enters the convolutional module, the following process is carried out: First, the convolutional module receives an input tensor, and then performs a convolution operation on the input tensor to extract local features and obtain an output feature map. Assume the input is , where C is the number of channels, and H and W are the width and height of the features; the convolutional kernel is , where and are the height and width of the convolutional kernel, then the output Y of the convolution operation is: ; where i and j are the positions of the output feature map.

[0040] Then, perform two-dimensional instance normalization on the convolved features. For the output Y of the convolution operation, calculate the mean μ and variance of each channel: ; ; The normalized output Z is: ; where is a very small value for numerical stability.

[0041] After the original Mel spectrogram is subjected to feature extraction through the convolutional module, the output result of the convolutional module is input into the attention module. The attention module can capture the global information of the audio signal and the internal long-distance connections within the features simultaneously from the time and frequency dimensions, thereby better modeling the potential characteristics of blade faults. The processing process of the attention module is as follows: First, the input feature Z is mapped to queries, keys, and values through a linear transformation, i.e.: 、 、 , where 、 、 are learnable weight matrices; Next, the attention weights are calculated by taking the dot product of the query and the key and normalizing, and the formula is: ; where is the dimension of the key, and softmax is used to normalize the weights into a probability distribution; Finally, the weighted sum of the attention weights and the values is output, that is, the final feature representation is obtained.

[0042] After the multi-layer perceptron flattens the final feature representation obtained above, it performs a non-linear transformation through a fully connected layer, thereby outputting an audio representation vector, which is the first potential audio representation in this step.

[0043] S4. Based on the first potential audio representation, use the decoder to obtain the reconstructed Mel spectrogram; In this step, the decoder includes a multi-layer perceptron, which restores the first potential audio representation generated by the first encoder to a high-dimensional Mel spectrogram feature to obtain the reconstructed Mel spectrogram.

[0044] S5. Based on the reconstructed Mel spectrogram, use the second encoder to obtain the second potential audio representation; In this step, the structure of the second encoder is the same as that of the first encoder. By further feature extraction of the reconstructed Mel spectrogram, a representation vector with richer feature expressions is obtained, which is the second potential audio representation in this step.

[0045] S6. According to the first potential audio representation and the second potential audio representation, calculate the anomaly score. If the anomaly score is greater than the preset threshold, it is determined that the blade of the wind power equipment is abnormal.

[0046] In this step, the calculation of the abnormal situation score is based on the difference between the first latent audio representation and the second latent audio representation. Since the model has learned the audio feature distribution in the normal working state of the wind power equipment blade, if the input original audio is the audio under normal conditions, it conforms to the audio feature distribution learned by the model, and the representation of the reconstructed audio (the second latent audio representation) is relatively close to the representation of the original audio (the first latent audio representation). On the contrary, if the input original audio is the audio under faulty conditions, the representation of the reconstructed audio (the second latent audio representation) is quite different from the representation of the original audio (the first latent audio representation). Therefore, by calculating the Euclidean distance between the two audio representations and comparing it with a set threshold, it can be determined whether the input audio is the audio generated during normal operation.

[0047] The calculation formula for the abnormal situation score is as follows: ; Where, is the abnormal situation score of the x-th audio, represents the first latent audio representation of the x-th audio, represents the second latent audio representation of the x-th audio, represents the L2 norm.

[0048] If the abnormal situation score is greater than the preset threshold, it is determined that the wind power equipment blade is abnormal.

[0049] Embodiment To verify the detection accuracy of the solution of the present invention for the faults of wind power equipment blades, in this embodiment, the detection accuracies of three different detection models for wind power equipment blades will be compared. The tested audio data includes hundreds of audio data collected from different parts of different wind power equipment blades in normal or abnormal states. The detection models used include the model proposed by the present invention, the CNN (Convolutional Neural Network) model in the prior art, and the CRNN (Convolutional Recurrent Neural Network) model.

[0050] Table 1 shows the accuracy rates of detecting the audio collected from three different parts on three different wind power equipment f1, f2, and f3 using the solution of the present invention.

[0051] Table 1 Accuracy Rates Detected by the Solution of the Present Invention

[0052] Table 2 shows the accuracy rates of detecting the audio collected from three different parts on three different wind power equipment f1, f2, and f3 using the CNN model.

[0053] Table 2 Accuracy Rates Detected by the CNN Model

[0054] Table 3 shows the accuracy rates of audio detection collected from three different parts on three different wind power devices f1, f2, and f3 using the CRNN model.

[0055] Accuracy Rates of Detection by CRNN Model

[0056] According to the above Tables 1 - 3, combined with Figure 3 , it can be seen that the average accuracy rates of the method proposed in the present invention on devices f1 and f2 are 78.82% and 77.98% respectively, slightly higher than those of the CNN and CRNN models; on device f3, the average accuracy rate of the method of the present invention reaches 80.20%, significantly higher than 70.10% of CNN and 70.29% of CRNN. Specifically for each acquisition part, the accuracy rate of the method of the present invention is higher than that of the other two models in most parts. For example, in part 1 of device f3, the accuracy rate of the method of the present invention is 81.11%, while that of CNN is 69.93% and that of CRNN is 70.53%. This indicates that the method proposed in the present invention has better accuracy and reliability in the fault detection of wind power device blades, can more effectively identify the fault conditions of the blades, and provides more powerful support for the maintenance and management of wind power devices.

[0057] Finally, it should be noted that the above embodiments are only preferred embodiments and do not limit the present invention. It should be pointed out that for those of ordinary skill in the art in the technical field, without departing from the spirit and scope protected by the claims of the present invention, several modifications, equivalent replacements, improvements, etc. can be made, and all should be included in the protection scope of the present invention.

Claims

1. A wind turbine blade fault detection method based on audio potential representation uses a trained fault detection model to perform fault detection on wind turbine blades, which is characterized by: The fault detection model includes a generator network using an encoder-decoder-encoder architecture, wherein the generator network includes a first encoder, a decoder, and a second encoder; and the detection method includes the following steps: S1. Collect the original audio of wind turbine blades in operation; S2. preprocess the original audio and extract Mel spectrum features to obtain the original Mel spectrum; S3. Based on the original Mel spectrum, using the first encoder to obtain a first potential audio representation; S4. Based on the first potential audio representation, using a decoder to obtain a reconstructed Mel spectrum; S5. Based on the reconstructed Mel spectrum, using a second encoder to obtain a second potential audio representation; S6. Calculate an abnormality score based on the first potential audio representation and the second potential audio representation. If the abnormality score is greater than a preset threshold, determine that the blade of the wind turbine equipment is abnormal.

2. The wind turbine blade fault detection method based on audio potential characterization according to claim 1, characterized in that: In step S1, collecting the original audio of the wind turbine blades in operation includes: Audio collection devices are arranged at different positions of the wind power equipment, and audio data of the wind power equipment blades when working is collected through the audio collection devices.

3. The wind turbine blade fault detection method based on audio potential characterization according to claim 1, characterized in that: In step S2, the original audio is preprocessed and the Mel spectrum features are extracted to obtain the original Mel spectrum, including: Perform pre-emphasis processing on the original audio signal; Divide the pre-emphasized audio signal into frame signals with a length of 20-40 milliseconds; Apply window function to each frame signal for processing; Perform fast Fourier transform on each frame signal processed by the window function to obtain a short-time spectrum; The short-time spectrum is filtered using a Mel filter bank, the linear frequency is converted into Mel frequency, and the original Mel spectrum is obtained.

4. The wind turbine blade fault detection method based on audio potential characterization according to claim 1, characterized in that: The first encoder and the second encoder both include three cascaded convolution modules, an attention module and a multi-layer perceptron; the convolution module is used to perform a convolution operation on the input Mel spectrum and perform two-dimensional instance normalization processing to obtain a normalized feature tensor; the attention module is used to calculate the attention weight on the normalized feature tensor to obtain a weighted feature representation; the multi-layer perceptron is used to flatten the weighted feature representation, perform a nonlinear transformation through a fully connected layer, and output an audio representation vector.

5. The wind turbine blade fault detection method based on audio potential characterization according to claim 1, characterized in that: The decoder includes a multi-layer perceptron, which is used to restore the first potential audio representation generated by the first encoder into a high-dimensional Mel spectrum feature to obtain a reconstructed Mel spectrum.

6. The wind turbine blade fault detection method based on audio potential characterization according to claim 1, characterized in that: In step S6, calculating the abnormality score according to the first potential audio representation and the second potential audio representation includes: ； in, Score the abnormality of the xth audio. represents the first potential audio representation of the xth audio, represents the second potential audio representation of the xth audio, represents the L2 norm.

7. The wind turbine blade fault detection method based on audio potential characterization according to claim 1, characterized in that: The fault detection model is trained in an offline manner, and a discriminator is introduced to form an adversarial network. The training process includes: a. Collect the original audio of wind turbine blades at work; b. Preprocess the original audio and extract the Mel spectrum features to obtain the original Mel spectrum; c. Based on the original Mel spectrum, the generator network in the fault detection model is used to obtain the reconstructed Mel spectrum; d. Add random noise to the original Mel spectrum to obtain the generated original Mel spectrum; e. Based on the generated original Mel spectrum, the generator network in the fault detection model is used to obtain the generated reconstructed Mel spectrum; f. Using the original Mel spectrum, the reconstructed Mel spectrum, the generated original Mel spectrum and the generated reconstructed Mel spectrum as inputs of the discriminator, and performing adversarial training with the generator network; g. Repeat steps af until the training termination condition is reached to obtain a trained fault detection model.

8. The wind turbine blade fault detection method based on audio potential characterization according to claim 7, characterized in that: In step f, during adversarial training, the loss function used by the fault detection model includes: ； in, is the total loss; , , They are the first part of the loss, the second part of the loss and the third part of the loss; are the weights corresponding to the first part of the loss, the second part of the loss, and the third part of the loss respectively; ； ； ； in, represents the original Mel spectrum or the generated original Mel spectrum; represents the reconstructed mel spectrum or the generated reconstructed mel spectrum; A first potential audio representation representing an original Mel spectrum or a generated original Mel spectrum obtained after being processed by a first encoder; A second potential audio representation obtained by processing the reconstructed mel-spectrogram or the generated reconstructed mel-spectrogram through a second encoder; represents the implicit function of the discriminator; Represents the Sigmoid function; represents the L1 norm; represents the L2 norm.

9. The wind turbine blade fault detection method based on audio potential characterization according to claim 8, characterized in that: In step f, during adversarial training, the Adam algorithm is used to update the parameters of the fault detection model. The updating process includes: ； ； ； ； ； ； in, represents the time step gradient at time t; It means to find the gradient; The network parameters are loss of time; is the exponential moving average; is the first-order decay rate of the moment estimate; is the corrected gradient mean at time t; represents the second-order moment at time step t-1; is the second-order decay rate of the moment estimate; represents the corrected squared gradient; represents the bias-corrected second-order moment gradient estimate; is a constant; is the learning rate.

10. Wind turbine blade fault detection system based on audio potential representation, It is characterized in that Including: sound collection equipment, wireless transmission module and intelligent terminal; The sound collection device is used to collect the original audio when the blades of the wind power equipment are working; The wireless transmission module is used to transmit the collected original audio to the intelligent terminal; The intelligent terminal is used to perform fault detection on a wind turbine blade according to the wind turbine blade fault detection method based on audio potential characterization as described in any one of claims 1 to 9.

Images