A Deep Learning-Based Acoustic Detection Method and System for Wind Turbine Blade Faults

Abstract

This invention discloses a deep learning-based acoustic detection method and system for wind turbine blade faults. The method first performs beamforming, high-pass filtering enhancement, and segmentation and augmentation on microphone array training data; then, it extracts short-time Fourier transform features, Mel filter bank features, short-time graphical Fourier transform features, and graphical Mel filter bank features from the training data; finally, it trains a wind turbine blade fault detection neural network; actual microphone array data undergoes beamforming, high-pass filtering enhancement, and signal feature extraction, and is input into the trained neural network to obtain the fault detection results. The system includes a server and a client; the client is implemented via a web page to adapt to user devices with different hardware and software, and communication between the client and server is achieved via the SSH port. This invention demonstrates high accuracy on a test set constructed from recorded wind turbine blade radiated noise data, and the system is adaptable to user devices with different hardware and software.

Description

Technical Field

[0001] This invention relates to the field of wind turbine blade fault detection technology, and in particular to a deep learning-based acoustic detection method and system for wind turbine blade faults. Background Technology

[0002] As a key component of wind turbines, blades account for 20% of the cost of a single wind turbine unit. Prolonged operation of wind turbines, coupled with environmental factors such as wind pressure, wind speed, and temperature, can lead to various blade failures, increasing maintenance costs and even affecting the performance and lifespan of the wind turbine. Monitoring the health status of large wind turbine blades, promptly and accurately detecting damage and providing early warnings, while minimizing downtime and maintenance costs, has become a key research focus in the wind power field. Since wind turbines generate significant noise during operation, some of which directly affects blade health, wind turbine blade damage detection methods based on wind turbine blade radiated noise signal processing have gradually developed in recent years. This method offers advantages such as high efficiency, harmlessness, convenient installation, and low cost. However, the collected blade radiated noise signals are often interfered with by strong background noise in wind farms, making the acoustic characteristics of faulty blades increasingly less obvious, thus reducing the accuracy of fault detection algorithms. Overcoming the harsh noise environment of wind farms and accurately extracting and identifying the acoustic signal characteristics of faulty blades are two important issues. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a method and system for wind turbine blade fault detection based on deep learning, thereby enabling fault detection of wind turbine blades.

[0004] The technical solution adopted in this invention is: an acoustic detection method for wind turbine blade faults based on deep learning, comprising the following steps:

[0005] Step 1: Acquire microphone array training data, and perform beamforming and high-pass filtering enhancement on the microphone array training data; obtain wind turbine blade radiated noise training data.

[0006] Step 2: The training data on radiated noise from the wind turbine blades were then divided and augmented;

[0007] Step 3: Extract signal features from the training data obtained in Step 2, including Short Time Fourier Transform (STFT), Mel Filter Bank (Fbank), Short Time Graph Fourier Transform (STGFT), and Graph Domain Mel Filter Bank (Gbank).

[0008] Step 4: Train the neural network for wind turbine blade fault detection;

[0009] Step 5: Perform beamforming and high-pass filtering enhancement on the actual microphone array data, extract signal features, and input them into the trained wind turbine blade fault detection neural network to obtain fault detection results.

[0010] Furthermore, step 1 specifically includes the following sub-steps:

[0011] Step 11, perform beamforming enhancement on the microphone array training data:

[0012]

[0013] Where Y is the discrete Fourier transform result of the enhanced signal, X m ω is the result of the discrete Fourier transform of the received signal of the m-th array element. m (f k ) is the beam weight for conventional beamforming of the array, f k It is the signal frequency corresponding to the k-th frequency point;

[0014] Step 12: Perform high-pass filtering on the enhanced single-channel signal.

[0015] Furthermore, step 2 specifically includes the following sub-steps:

[0016] Step 21: Divide the training data for wind turbine blade radiated noise into durations of 10s-20s, ensuring the duration remains consistent throughout the training process.

[0017] Step 22: The enhanced single-channel signal is augmented with data including random pitch shift, random velocity variation, random energy gain, superimposed colored noise, superimposed background wind noise, and random low-pass filter.

[0018] Furthermore, step 3 is specifically as follows:

[0019] Step 31: Extract the Short-Time Fourier Transform (STFT) feature of the input signal. First, the input signal s is divided into frames to obtain the framed result.

[0020]

[0021] 0≤t<T, h=t×N s

[0022] Where N f N is the frame length. sThe frame shift is represented by T, where T is the total number of frames after framing. A window function is used after framing to avoid spectral leakage. The windowed signal is then converted from the time domain to the frequency domain by an FFT operation. By stacking the FFT calculation results of all signals after framing, the Short-Time Fourier Transform (STFT) characteristic of the signal can be obtained.

[0023]

[0024] The calculation process of FFT for a single frame signal is shown below, where N fft Here, w(n) represents the number of points in the Fourier transform, and w(n) represents the window function used.

[0025]

[0026] Step 32: Extract the Mel filter bank features (Fbank) of the input signal; the Mel filter bank features (Fbank) of the signal are the STFT energy spectrum of the signal, obtained by passing the signal through a set of Mel filter banks H. m The result is that the energy spectrum is passed through a set of Mel filters, which are triangular filters. The characteristic Fbank of the Mel filter bank, expressed in dB, is the signal, as shown in the following formula:

[0027]

[0028] Step 33: Extract the short-time graph Fourier transform (STGFT) feature of the input signal; divide the input signal s into frames. The signal after framing is regarded as a stationary signal within the frame. Perform graph Fourier transform on each frame signal according to the designed signal graph structure. Then, stack the GFT calculation results of all signals after framing to obtain the short-time graph Fourier transform feature (STGFT).

[0029]

[0030] Where U is the eigenvector matrix obtained by eigenvalue decomposition of the Laplacian matrix L of the graph, L = D – A;

[0031]

[0032] Step 34: Extract the graph domain Mel filter bank features Gbank of the input signal; after obtaining the short-time graph Fourier transform features STGFT of the signal, square the short-time graph Fourier transform features STGFT to obtain the energy spectrum in the graph domain, and then pass the energy spectrum through a set of graph domain Mel filters Gbank. m When expressed in dB, this represents the graphic domain Mel filter bank characteristic of the signal:

[0033]

[0034] Furthermore, step 4 is specifically as follows:

[0035] The extracted features are input into a deep neural network for training. The network architecture can be ResNet, ECAPA-TDNN, or MFA-Conformer. The input features are normalized through learnable layers for adaptive scaling before being fed into the neural network. Finally, the estimated predicted value is obtained after passing through a fully connected layer (FC) and the SoftMax activation function. The maximum value corresponds to either category 0 or 1, which is either normal or faulty, and the number of neurons in the fully connected layer is N1.

[0036] The fault detection problem is solved using a classification model, with normal and faulty categories each. Cross-entropy is used as the loss function, and the calculation formula is as follows:

[0037]

[0038] Where p(x) n ) represents the true probability of belonging to the nth category. This represents the probability of predicting that the value belongs to the nth category.

[0039] On the other hand, this invention also provides a deep learning-based acoustic detection system for wind turbine blade faults. The system includes a server and a client. The client is implemented based on a web page to adapt to user devices with different software and hardware. Communication between the client and the server is implemented based on the SSH port.

[0040] The server is responsible for storing and processing data, and includes a data storage module, a data reading module, and a data processing module; the system is divided into online fault detection function and historical data analysis function according to the reading method of the data reading module.

[0041] The data storage module is responsible for storing the key data required for system operation. The stored data includes two parts: historical wind turbine blade radiation noise data collected by the microphone array and trained neural network weight data. The historical wind turbine blade radiation noise data collected by the microphone array is tagged when stored, including the region and wind turbine number.

[0042] The data reading module offers two reading methods: online fault detection and historical data analysis. For online fault detection, the server receives array data uploaded by the user from the client, and then saves and reads the data. For historical data analysis, the system does not require the user client to collect data or store data to be processed. The user only needs to select the data stored on the central server based on the region and turbine number corresponding to the wind turbine blade radiated noise data. In this case, the data reading module directly reads the corresponding data from the storage space. These data reading methods correspond to the system's two main functions: online fault detection and historical data analysis.

[0043] The data processing module first accepts user input requests from the user interaction module, including the signal features and neural network model selected by the user, and loads the corresponding signal feature operation functions and neural network weights. Then, the data processing module receives the array signal read from the data reading module and performs calculations on it, including a process of "preprocessing, feature extraction, and neural network discrimination". The preprocessing includes dividing and augmenting the wind turbine blade radiated noise training data. The feature extraction includes extracting short-time Fourier transform features, Mel filter bank features, short-time graph Fourier transform features, and graph domain Mel filter bank features from the obtained training data. The wind turbine blade fault detection neural network is trained. The actual microphone array data is enhanced by beamforming and high-pass filtering, and signal features are extracted and input into the trained wind turbine blade fault detection neural network to obtain fault detection results. After the calculation is completed, the module returns the waveform drawing results before and after signal preprocessing, the signal feature visualization results, and the fault detection results, and then saves the calculation results for use by the user interaction module.

[0044] The client is responsible for interacting with the user and includes a user management module and a user interaction module;

[0045] The user management module verifies user identity through account and password verification and allows them to access specific resources and functions within the system to ensure system security and controllable permissions. When a user requests access to the wind turbine blade fault detection system, they need to enter an account and password for verification. If the verification information is correct, the system will authorize the user to access resources and functions within their specific permissions. If the user enters incorrect verification information, the system will refuse access.

[0046] The user interaction module includes two aspects: the system user input interface and the calculation result display interface. The user input interface is responsible for accepting user input requests and sending them to the server. In this paper, the user input interface includes a data upload interface, a data selection interface, a signal feature selection interface, and a neural network model selection interface. The calculation result display interface is responsible for displaying the calculation results returned by the user request to the user.

[0047] Furthermore, the server and client need to communicate within the same local area network.

[0048] Furthermore, the server is implemented using the Python programming language and the PyTorch deep learning framework.

[0049] The beneficial effects of this invention are as follows: A spatiotemporal joint preprocessing method is used to enhance the received microphone array signal, thereby suppressing complex background noise and interference in wind farms. For the enhanced signal, this invention extends traditional frequency domain signal feature analysis to the graph domain based on traditional STFT and Fbank features in the frequency domain, combined with graph signal processing theory. Based on signal feature extraction methods and a joint deep learning fault detection algorithm, a wind turbine blade fault detection system is designed, which is adaptable to user equipment with different hardware and software devices. Attached Figure Description

[0050] Figure 1 This is a flowchart of the wind turbine blade fault detection algorithm in a specific embodiment of the present invention;

[0051] Figure 2 This is a system workflow diagram in a specific embodiment of the present invention;

[0052] Figure 3 This is a block diagram of a wind turbine blade fault detection system module in a specific embodiment of the present invention. Detailed Implementation

[0053] The acoustic detection system for wind turbine blade faults proposed in this invention will be described in detail below with reference to the embodiments and accompanying drawings.

[0054] Figure 1 A flowchart of the wind turbine blade fault detection algorithm used in the system is presented. The algorithm consists of two phases: training and inference. In the training phase, the data in the dataset is first preprocessed and augmented. Then, the dataset size is expanded through data augmentation. Short-time features of the signal are selected and input into the neural network, which is then trained using the backpropagation algorithm. In the inference phase, the selected features and the trained neural network weights are loaded. The received array signal is then preprocessed, features are extracted, and the neural network makes predictions to obtain the fault detection result.

[0055] The specific implementation of the technical method of the present invention is as follows:

[0056] Step 1: Acquire microphone array training data, and perform beamforming and high-pass filtering enhancement on the microphone array training data; obtain wind turbine blade radiated noise training data, including the following sub-steps:

[0057] Step 11: Beamforming enhancement is performed on the signal received by the microphone array.

[0058]

[0059] Where Y is the discrete Fourier transform result of the enhanced signal, X m ω is the result of the discrete Fourier transform of the received signal of the m-th array element. m (f k ) is the beam weight for conventional beamforming of the array, f k It is the signal frequency corresponding to the k-th frequency point;

[0060] Step 12: Perform high-pass filtering on the enhanced single-channel signal, with a cutoff frequency f. l =200Hz.

[0061] Step 2 involves dividing and augmenting the wind turbine blade radiated noise training data, including the following sub-steps:

[0062] Step 21: The wind turbine blade radiated noise signals in the dataset come from wind farms in Chongming (Shanghai), Shuozhou (Shanxi), Pan'an (Zhejiang), Suzhou (Anhui), and Sheyang (Jiangsu). The training / test data duration is 15 seconds, the sampling rate is 20 kHz, the training set contains 23,940 data points, and the test set contains 4,430 data points. Data from the Jiangsu Sheyang wind farm only appears in the test set.

[0063] Step 22: Augment the training data in the dataset. The augmented single-channel signal is augmented with random pitch shift, random velocity change, random energy gain, superimposed colored noise, superimposed background wind noise, and random low-pass filter.

[0064] Step 3: Extract signal features from the training data obtained in Step 2, including Short-Time Fourier Transform (STFT), Mel-Filter Bank (Fbank) features, Short-Time Graph Fourier Transform (STGFT) features, and Graph Domain Mel-Filter Bank (Gbank) features. This includes the following sub-steps:

[0065] Step 31: Extract the STFT features of the input signal. First, the long input signal s is divided into frames:

[0066]

[0067] 0≤t<T, h=t×N s

[0068] Wherein, N f N is the frame length. sThe frame shift is represented by T, where T is the total number of frames after framing. A window function is then used to avoid spectral leakage. The windowed signal is then converted from the time domain to the frequency domain using an FFT operation. The STFT result of the signal is obtained by stacking the FFT calculation results of all the signals after framing.

[0069]

[0070] The calculation process of FFT for a single frame signal is shown below, where N fft is the number of points in the Fourier transform; w(n) is the window function used;

[0071]

[0072] Step 32: Extract the Fbank features of the input signal. The Fbank features of the signal are the result of passing the STFT energy spectrum of the signal through a set of Mel filter banks. The energy spectrum is then passed through a set of Mel filter H... m The result, expressed in dB, is the Fbank characteristic of the signal. The Mel filter bank is a 40th-order triangular filter in the frequency range of 0–10 kHz;

[0073]

[0074] Step 33: Extract the short-time graph Fourier transform (STGFT) feature of the input signal; divide the input signal s into frames. The signal after framing is regarded as a stationary signal within the frame. Perform graph Fourier transform on each frame signal according to the designed signal graph structure. Then, stack the GFT calculation results of all signals after framing to obtain the short-time graph Fourier transform feature (STGFT).

[0075]

[0076] Where U is the eigenvector matrix obtained by eigenvalue decomposition of the Laplacian matrix L of the graph, L = D – A;

[0077]

[0078] Step 34: Extract the graph domain Mel filter bank features Gbank of the input signal; after obtaining the short-time graph Fourier transform features STGFT of the signal, square the short-time graph Fourier transform features STGFT to obtain the energy spectrum in the graph domain, and then pass the energy spectrum through a set of graph domain Mel filters Gbank. m When expressed in dB, this represents the graphic domain Mel filter bank characteristic of the signal:

[0079]

[0080] Step 4: Train the neural network for wind turbine blade fault detection.

[0081] The extracted features are input into a deep neural network for training. The network architecture can be ResNet, ECAPA-TDNN, or MFA-Conformer. The input features are normalized through learnable layers for adaptive scaling before being fed into the neural network. Finally, the estimated predicted value is obtained after passing through a fully connected layer (FC) and the SoftMax activation function. The maximum value corresponds to either category 0 or 1, which is either normal or faulty, and the number of neurons in the fully connected layer is N1.

[0082] The fault detection problem is solved using a classification model, with normal and faulty categories each. Cross-entropy is used as the loss function, and the calculation formula is as follows:

[0083]

[0084] Where p(x) n ) represents the true probability of belonging to the nth category. This represents the probability of predicting that the value belongs to the nth category.

[0085] Step 5: Perform beamforming and high-pass filtering enhancement on the actual microphone array data, extract signal features, and input them into the trained wind turbine blade fault detection neural network to obtain fault detection results.

[0086] According to a second aspect of the present invention, a deep learning-based acoustic detection system for wind turbine blade faults is provided, such as... Figure 2 The diagram illustrates the workflow of a wind turbine blade fault detection system. The system collects radiated noise from each wind turbine generator in the wind farm using a microphone array and transmits this data to a central control system for storage and processing. After data is collected and uploaded to the central server, online fault detection results are obtained. The system consists of two main parts: a server and a client, corresponding to the central server and the user, respectively. The server and client communicate via a network. The server is responsible for storing and calculating the collected wind turbine blade radiated noise and responding to user requests. The client requires a user interface, allowing users to send requests to view the wind turbine blade radiated noise stored on the server and its calculation results. The client portion of the system is compatible with various devices, operating systems, and other hardware and software to meet the needs of different users in the wind farm. This ensures that users can access the system client regardless of their hardware or software configuration, as long as they have access to the communication network.

[0087] like Figure 3As shown, the entire wind turbine blade fault detection system consists of two main parts: a server and a client. The server and client need to communicate within the same local area network. The server is implemented using the Python programming language and the PyTorch deep learning framework, while the client is implemented using a web page to adapt to different user devices with varying hardware and software. Communication between the client and server is achieved via the SSH port.

[0088] The server is responsible for storing and processing data, including a data storage module, a data reading module, and a data processing module. Based on the data reading method, it can be divided into two functions: online fault detection and historical data analysis.

[0089] (1) Data storage module: The data storage module is responsible for storing the key data required for system operation. The stored data mainly consists of two parts: historical wind turbine blade radiation noise data collected by the microphone array and trained neural network weight data. The historical wind turbine blade radiation noise data collected by the microphone array is tagged during storage, including the region and wind turbine number. Since the wind turbine blade radiation noise data used in this paper comes from multiple wind farms, the region of the data stored in this paper is the name of the wind farm. When actually used in a single wind farm, it can be replaced with a general area divided by direction, such as the east or south of the wind farm.

[0090] (2) Data Reading Module: The data reading module has two reading methods. For the first method, the server receives array data uploaded by the user from the client and saves and reads the data. For the second method, the system does not require the user client to collect data or store data to be processed. The user is only required to select the data stored on the central server based on the region and turbine number corresponding to the wind turbine blade radiated noise data. In this case, the data reading module directly reads the corresponding data from the storage space. These data reading methods correspond to the system's two main functions: online fault detection and historical data analysis.

[0091] (3) Data Processing Module: The data processing module first accepts user input requests from the user interaction module, including the signal features and neural network model selected by the user, and loads the corresponding signal feature operation functions and neural network weights. Then, the data processing module receives the array signal read from the data reading module and performs calculations on it, including the processes of "preprocessing, feature extraction, and neural network discrimination". After the calculation is completed, the module returns the waveform drawing results before and after signal preprocessing, the signal feature visualization results, and the fault detection results, and then saves the calculation results for use by the user interaction module. The signal features available to the user in this system are STFT, Fbank, STGFT, Gbank, and Fbank+Gbank, and the neural network structures available to the user are ResNet, ECAPA-TDNN, and MFA-Conformer.

[0092] The client is responsible for interacting with users and includes a user management module and a user interaction module.

[0093] (1) User Management Module: User management is the system's method of verifying user identity through account and password authentication, and allowing them to access specific resources and functions within the system to ensure system security and controllable permissions. When a user requests access to the wind turbine blade fault detection system, they need to enter their account and password for verification. If the verification information is correct, the system will authorize the user to access resources and functions within their specific permissions; if the user enters incorrect verification information, the system will refuse access.

[0094] (2) User Interaction Module: User interaction comprises two aspects: the system user input interface and the calculation result display interface. The user input interface is responsible for accepting user input requests and sending them to the server. In this paper, the user input interface mainly includes the data upload interface, data selection interface, signal feature selection, and neural network model selection. The calculation result display interface is responsible for displaying the calculation results returned by the user's request to the user. In this paper, the calculation result display interface is mainly responsible for waveform drawing before and after signal preprocessing, signal feature visualization, and fault detection log display.

[0095] Example:

[0096] This example uses a 4×4 16-element rectangular microphone array with an element spacing of 5cm. When collecting radiated noise signals from wind turbine blades, the array is placed directly below the wind turbine and connected to the industrial control chassis to complete data acquisition. The signal sampling frequency is 20kHz, and short-time feature extraction is performed using a 512-point window and a 256-point frame shift. Input features are selected as STFT, Fbank, STGFT, Gbank, and Fbank+Gbank concatenation features, respectively. Network structures include ResNet, ECAPA-TDNN, and MFA-Conformer, and corresponding neural networks are constructed. The neural networks are iteratively trained, and a fault detection system is built based on the trained neural networks.

[0097] By opening a service port on the central server, users can access the wind turbine fault detection system through a browser and enter the system homepage. Using a macOS-powered Macbook as the test device, users can log in by entering their username and password after accessing the homepage.

[0098] The wind turbine blade fault detection system integrated into the system was tested using a constructed test set. The test set contained 4430 data points from wind farms in Chongming (Shanghai), Shuozhou (Shanxi), Pan'an (Zhejiang), Suzhou (Anhui), and Sheyang (Jiangsu), with data from the Sheyang wind farm appearing only in the test set. The proposed algorithm, combining Fbank+Gbank splicing features with an MFA-Conformer network, achieved a fault detection accuracy of 95.5% on the test set.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based acoustic detection method for wind turbine blade faults, characterized in that, Includes the following steps: Step 1: Acquire microphone array training data, and perform beamforming and high-pass filtering enhancement on the microphone array training data to obtain wind turbine blade radiated noise training data; Step 1 specifically includes the following sub-steps: Step 11, perform beamforming enhancement on the microphone array training data: ； in It is the result of the discrete Fourier transform of the enhanced signal. It is the first Discrete Fourier transform results of the received signals of each array element It is the beam weight of the array conventional beamforming. It is the signal frequency corresponding to the k-th frequency point; Step 12: Perform high-pass filtering on the enhanced single-channel signal; Step 2: Divide and augment the training data of wind turbine blade radiated noise; Step 3: Extract signal features from the training data obtained in Step 2, including Short Time Fourier Transform (STFT), Mel Filter Bank (Fbank), Short Time Graph Fourier Transform (STGFT), and Graph Domain Mel Filter Bank (Gbank). Step 3 is described in detail below: Step 31: Extract the Short-Time Fourier Transform (STFT) feature of the input signal. First, analyze the input signal... Perform frame segmentation to obtain the framed result. : ； in For frame length, The frame shift is represented by T, where T is the total number of frames after framing. A window function is used after framing to avoid spectral leakage. The windowed signal is then converted from the time domain to the frequency domain by an FFT operation. By stacking the FFT calculation results of all signals after framing, the Short-Time Fourier Transform (STFT) characteristic of the signal can be obtained. ； The calculation process of FFT for a single frame signal is shown below, where, The number of points in the Fourier transform. The window function used; ； Step 32: Extract the Mel filter bank features (Fbank) of the input signal; the Mel filter bank features (Fbank) of the signal are the STFT energy spectrum of the signal after passing through a set of Mel filters. The result is that the energy spectrum is passed through a set of Mel filters, which are triangular filters. The characteristic Fbank of the Mel filter bank, expressed in dB, is the signal's characteristic Fbank, as shown in the following formula: ； Step 33: Extract the short-time Fourier transform (STGFT) features of the input signal; for the input signal... The signal is divided into frames, and the signal after framing is considered as a stationary signal within the frame. According to the designed signal graph structure, the graph Fourier transform is performed on the signal of each frame. Then, the GFT calculation results of all signals after framing are stacked to obtain the short-time graph Fourier transform feature STGFT. ； Where U is the eigenvector matrix obtained by eigenvalue decomposition of the Laplacian matrix L of the graph, L = D – A; ； Step 34: Extract the graph domain Mel filter bank features Gbank of the input signal; after obtaining the short-time graph Fourier transform features STGFT of the signal, square the short-time graph Fourier transform features STGFT to obtain the energy spectrum in the graph domain, and then pass the energy spectrum through a set of graph domain Mel filters. When expressed in dB, this represents the graphic domain Mel filter bank characteristic of the signal: ； Step 4: Train the neural network for wind turbine blade fault detection; Step 5: Perform beamforming and high-pass filtering enhancement on the actual microphone array data, extract signal features, and input them into the trained wind turbine blade fault detection neural network to obtain fault detection results.

2. The acoustic detection method for wind turbine blade faults based on deep learning according to claim 1, characterized in that, Step 2 specifically includes the following sub-steps: Step 21: Divide the training data for wind turbine blade radiated noise into durations of 10s-20s, ensuring the duration remains consistent throughout the training process. Step 22: The enhanced single-channel signal is augmented with data including random pitch shift, random velocity variation, random energy gain, superimposed colored noise, superimposed background wind noise, and random low-pass filtering.

3. The acoustic detection method for wind turbine blade faults based on deep learning according to claim 1, characterized in that, Step 4 is as follows: The extracted features are input into a deep neural network for training. The network architecture can be ResNet, ECAPA-TDNN, or MFA-Conformer. The input features are normalized through learnable layers for adaptive scaling before being fed into the neural network. Finally, the estimated predicted value is obtained after passing through a fully connected layer (FC) and the SoftMax activation function. , The maximum value corresponds to either category 0 or 1, representing normal or faulty, where the number of neurons in the fully connected layer is... ; The fault detection problem is solved using a classification model, with normal and faulty categories each. Cross-entropy is used as the loss function, and the calculation formula is as follows: ； in This represents the true probability of belonging to the nth category. This represents the probability of predicting that the value belongs to the nth category.

4. A deep learning-based acoustic detection system for wind turbine blade faults, used to implement the method described in any one of claims 1-3, characterized in that, The system consists of two parts: a server and a client. The client is implemented based on a web page to adapt to user devices with different software and hardware, and communication between the client and the server is implemented based on the SSH port. The server is responsible for storing and processing data, and includes a data storage module, a data retrieval module, and a data processing module. The system is divided into online fault detection function and historical data analysis function according to the reading method of the data reading module; The data storage module is responsible for storing the key data required for system operation. The stored data includes two parts: historical wind turbine blade radiation noise data collected by the microphone array and trained neural network weight data. The historical wind turbine blade radiation noise data collected by the microphone array is tagged when stored, including the region and wind turbine number. The data reading module offers two reading methods: online fault detection and historical data analysis. For online fault detection, the server receives array data uploaded by the user from the client, and then saves and reads the data. For historical data analysis, the system does not require the user client to collect data or store data to be processed. The user only needs to select the data stored on the central server based on the region and turbine number corresponding to the wind turbine blade radiated noise data. In this case, the data reading module directly reads the corresponding data from the storage space. These data reading methods correspond to the system's two main functions: online fault detection and historical data analysis. The data processing module first accepts user input requests from the user interaction module, including the signal features and neural network model selected by the user, and loads the corresponding signal feature operation functions and neural network weights. Then, the data processing module receives the array signal read from the data reading module and performs calculations on it, including the processes of "preprocessing, feature extraction, and neural network discrimination". The preprocessing includes dividing and augmenting the wind turbine blade radiated noise training data. The feature extraction includes extracting short-time Fourier transform features, Mel filter bank features, short-time graph Fourier transform features, and graph domain Mel filter bank features from the obtained training data. Training a neural network for wind turbine blade fault detection; The actual microphone array data is enhanced by beamforming and high-pass filtering, and the signal features are extracted and input into the trained wind turbine blade fault detection neural network to obtain the fault detection results. After the calculation is completed, the module returns the waveform drawing results before and after signal preprocessing, the signal feature visualization results, and the fault detection results. Then, the calculation results are saved for use by the user interaction module. The client is responsible for interacting with the user and includes a user management module and a user interaction module; The user management module verifies user identity through account and password verification and allows them to access specific resources and functions within the system to ensure system security and controllable permissions. When a user requests access to the wind turbine blade fault detection system, they need to enter an account and password for verification. If the verification information is correct, the system will authorize the user to access resources and functions within their specific permissions. If the user enters incorrect verification information, the system will refuse access. The user interaction module includes two aspects: the system user input interface and the calculation result display interface. The user input interface is responsible for accepting user input requests and sending them to the server. In this paper, the user input interface includes a data upload interface, a data selection interface, a signal feature selection interface, and a neural network model selection interface. The calculation result display interface is responsible for displaying the calculation results returned by the user request to the user.

5. The deep learning-based acoustic detection system for wind turbine blade faults according to claim 4, characterized in that, The server and client need to communicate within the same local area network.

6. The deep learning-based acoustic detection system for wind turbine blade faults according to claim 4, characterized in that, The server is implemented using the Python programming language and the PyTorch deep learning framework.

Images