A fan blade damage detection method and system based on a multi-frequency attention network

Abstract

The application provides a fan blade damage detection method and system based on a multi-frequency attention network, and belongs to the technical field of fan blade damage detection. The application adopts octave convolution extraction to extract features, and then decomposes the features into low-frequency groups and high-frequency groups. Then, the channel attention is used to recalibrate the features of the two parts respectively. The low-frequency group features represent the smooth changes of the pixels in the picture, so the low-frequency group features can help to determine whether there is a damage. The high-frequency group features represent the rapid changes of the pixels in the picture, so the high-frequency group features can help to focus on the edges of the damage, which will help to eliminate the influence of the natural background in the picture and find small damages. The method provided by the application can reduce the influence of the natural background, and significantly improve the discrimination accuracy.

Description

Technical Field

[0001] This application belongs to the field of wind turbine blade damage detection technology, specifically relating to a wind turbine blade damage detection method and system based on a multi-frequency attention network. Background Technology

[0002] Wind energy is a low-cost, environmentally friendly renewable energy source. However, due to the geographical constraints of wind resources, wind farms are generally located in remote areas such as the northwest, plateaus, coastlines, and offshore. Wind turbines operate in harsh environments with unstable loads. With increasing operating time, wind turbine blades are highly susceptible to various failures such as cracks, corrosion, burning, and lightning strikes, leading to abnormal turbine operation and high maintenance costs. Therefore, timely detection and analysis of blade damage can increase the stability of wind turbines, extend blade lifespan, and reduce downtime.

[0003] In existing technologies, research on wind turbine blade condition monitoring and fault diagnosis largely relies on sensors, such as vibration sensors, acoustic sensors, strain sensors, and ultrasonic sensors. However, sensor-based detection methods have the following problems in practical applications: 1) Installing sensors on wind turbines requires additional manpower and resources; 2) The reliability of power supply and the complexity of sensor maintenance need to be considered; 3) Ensuring the reliability of sensor signal transmission is crucial.

[0004] With the rise of drone technology, computer vision-based methods have offered new insights into this problem. However, accurately identifying blade images captured by drones remains a particularly challenging task. First, most early damage is small and sparsely distributed, yet it is crucial for operation and maintenance. Second, the background information of wind turbine blades is more complex, with varying sizes, shapes, and textures of blade damage, posing new challenges to wind turbine blade damage detection.

[0005] Therefore, it is necessary to provide a wind turbine blade damage detection method and system based on a multi-frequency attention network to solve the problems mentioned in the background art. Summary of the Invention

[0006] This application provides a wind turbine blade damage detection method and system based on a multi-frequency attention network. It directly collects image data of wind turbine blades and then uses machine learning to identify blade faults. This effectively solves the problem of low classification accuracy in the prior art due to the small scale, sparse distribution, complex background information of images, and many types of damage in the early stage of wind turbine blade damage.

[0007] To solve the above-mentioned technical problems, the technical solution of this application is as follows:

[0008] This application provides a wind turbine blade damage detection method based on a multi-frequency attention network, comprising the following steps:

[0009] Step S1: Collect historical data of wind turbine blades to form a dataset;

[0010] Step S2: Construct a convolutional neural network. Input the dataset into the convolutional neural network for training. The convolutional neural network has four layers. Each layer contains several modules composed of octave convolution and channel attention. The input is separated into high-frequency and low-frequency groups after entering the first layer of the network, and then merged in the fourth layer. The merged features are then fed into a fully connected layer to output the classification result. During training, any module first uses octave convolution to extract the input features and divides the extracted features into high-frequency and low-frequency groups. Then, channel attention is used to calibrate the features of the high-frequency and low-frequency groups. The calibrated two sets of features are then fed into the next module as input.

[0011] In the first three layers of the convolutional neural network, channel attention uses a 2D-DCT-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups. This utilizes more frequency domain implicit information to generalize channel attention to the high-frequency and low-frequency groups, thereby calibrating the feature representations of the high-frequency and low-frequency groups. In the last layer, channel attention uses a GAP-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups, thereby calibrating the feature representations of the high-frequency and low-frequency groups.

[0012] Step S3: For any wind turbine blade damage detection, input the collected wind turbine blade data into the trained convolutional neural network, and output the judgment result of wind turbine blade damage.

[0013] Preferably, the kernel of an octave convolution is divided into four parts, which are represented as follows:

[0014] (1) Convolution kernel W from high frequency group to high frequency group H→H ;

[0015] (2) Convolution kernel W from high frequency group to low frequency group H→L ;

[0016] (3) Convolution kernel W from low-frequency group to high-frequency group L→H ;

[0017] (4) Convolution kernel W from low-frequency group to low-frequency group L→L ;

[0018] Input X includes high-frequency input X H and low-frequency input X L The process of an octave convolutional neural network extracting features from an input image X in a dataset includes:

[0019] High-frequency feature extraction: High-frequency group input X H After convolution kernel W H→H The output Y is obtained after convolution. H→H Low-frequency group input X L First, it passes through the convolution kernel W L→H The convolution operation is followed by upsampling to expand the resolution to the same level as the high-frequency group output resolution, resulting in the output Y. L→H ;Y H→H and Y L→H After the addition operation, the high-frequency characteristic Y is output. H ;

[0020] Low-frequency feature extraction: low-frequency input X L After convolution kernel W L→L The output Y is obtained after convolution. L→L High-frequency input X H First, the resolution is reduced to the same level as the low-frequency input resolution through downsampling, and then it is processed by convolution kernel W. H→L The convolution operation yields the output Y. H→L ;Y L→L With Y H→L After the addition operation, the low-frequency characteristic Y is output. L .

[0021] Preferably, the transformation formula for 2D-DCT is as follows:

[0022]

[0023] s.tμ∈{0,1,…,A-1},v∈{0,1,…B-1};

[0024] in This represents the input to 2D-DCT. This represents the output of 2D-DCT, where A and B represent the height and width of the 2D-DCT input, respectively, and u and v represent the two-dimensional frequency components, respectively; i and j represent the pixel index, i.e., the pixel in the i-th row and j-th column.

[0025] Use weights To represent the basis functions of 2D-DCT:

[0026]

[0027] The calculation process of 2D-DCT is as follows:

[0028] N frequency components are selected to generalize channel attention. The 2D-DCT weights corresponding to the N / 2 higher frequencies among the N frequencies are assigned to the high-frequency group, and the 2D-DCT weights corresponding to the N / 2 lower frequencies among the N frequencies are assigned to the low-frequency group. Each channel in the high-frequency group and the low-frequency group is calculated using the assigned N / 2 2D-DCT weights to obtain one-dimensional feature representations of N / 2 different frequencies.

[0029] Preferably, the one-dimensional feature of the N / 2 frequency components of the l-th channel of the high-frequency group. Represented as:

[0030]

[0031] stk∈{1,2,…N / 2}, l∈{1,2,…(1-α)C};

[0032] In the formula, α represents the proportion of the total number of low-frequency channels to the total number of channels; l represents the channel number; and k represents the frequency number. The 2D-DCT weight of the pixel in the i-th row and j-th column within the k-th frequency component of the high-frequency group is represented by C; C represents the total number of channels.

[0033] One-dimensional feature corresponding to the k-th frequency component of all channels in the high-frequency group Represented as:

[0034]

[0035] In the formula, cat(·) represents a chain operation;

[0036] Freq, the final one-dimensional feature of the entire high-frequency group H Represented as:

[0037]

[0038] In the formula, Represents the fully connected layer coefficients of the high-frequency group;

[0039] One-dimensional feature of the N / 2 frequency components of the l-th channel of the low-frequency group It is expressed as follows:

[0040]

[0041] stk∈{1,2,…N / 2}, l∈{1,2,…αC};

[0042] One-dimensional feature corresponding to the k-th frequency component of all channels in the low-frequency group Represented as:

[0043]

[0044] The final one-dimensional feature Freq of the low-frequency group's full channels L It is expressed as follows:

[0045]

[0046] In the formula, Represents the fully connected layer coefficients of the high-frequency group;

[0047] Therefore, the one-dimensional feature of the entire channel is represented as:

[0048] Freq = cat([Freq H ,Freq L ]).

[0049] Preferably, the calculation process for GAP is as follows:

[0050]

[0051] Therefore, the one-dimensional feature of the entire channel is represented as:

[0052] Freq=cat([Freq1,Freq2,…Freq C ]).

[0053] This application also provides a wind turbine blade damage detection system based on a multi-frequency attention network, including:

[0054] The data acquisition module is used to collect historical data of wind turbine blades to form a dataset;

[0055] The network construction module is used to build a convolutional neural network. The dataset is input into the convolutional neural network for training. The convolutional neural network has four layers. Each layer contains several modules composed of octave convolution and channel attention. The input is separated into high-frequency and low-frequency groups after entering the first layer of the network, and then merged in the fourth layer. The merged features are then fed into a fully connected layer to output the classification result. During training, any module first uses octave convolution to extract the input features and divides the extracted features into high-frequency and low-frequency groups. Then, channel attention is used to calibrate the features of the high-frequency and low-frequency groups. The calibrated two sets of features are then fed into the next module as input.

[0056] In the first three layers of the convolutional neural network, channel attention uses a 2D-DCT-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups. This utilizes more frequency domain implicit information to generalize channel attention to the high-frequency and low-frequency groups, thereby calibrating the feature representations of the high-frequency and low-frequency groups. In the last layer, channel attention uses a GAP-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups, thereby calibrating the feature representations of the high-frequency and low-frequency groups.

[0057] The damage discrimination module, for any wind turbine blade damage detection, inputs the collected wind turbine blade data into the trained convolutional neural network and outputs the discrimination result of wind turbine blade damage.

[0058] The beneficial effects of this application are as follows:

[0059] This paper employs octave convolution to extract features and decomposes these features into low-frequency and high-frequency components. Then, a two-dimensional discrete cosine transform is used to recalibrate both components. Low-frequency features represent the smooth variations in pixel values ​​within an image, thus helping to roughly determine the presence of damage. High-frequency features represent the rapid variations in pixel values, helping to focus on the edges of damage. This is beneficial for eliminating the influence of natural background and detecting subtle damage. The method provided in this application reduces the influence of natural background and significantly improves the accuracy of the discrimination. Attached Figure Description

[0060] Figure 1 This diagram illustrates the network architecture of the convolutional neural network provided in this application.

[0061] Figure 2 A diagram showing the network architecture of an octave convolution;

[0062] Figure 3 A schematic diagram illustrating the selection of characteristic frequency components;

[0063] Figure 4 In the image, (a), (b), and (c) represent the original image, high-frequency feature map, and low-frequency feature map of the wind turbine blade, respectively.

[0064] Figure 5 A network architecture diagram illustrating the channel attention mechanism based on 2D-DCT;

[0065] Figure 6 This diagram illustrates the network architecture based on the GAP channel attention mechanism. Detailed Implementation

[0066] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0067] Please refer to the following: Figures 1-6 This application provides a wind turbine blade damage detection method based on a multi-frequency attention network, comprising the following steps:

[0068] Step S1: Collect historical data of wind turbine blades to form a dataset.

[0069] In naturally collected wind turbine blade images, the ratio of intact blade images to damaged blade images is often highly imbalanced, with intact blade images far outnumbering damaged ones. To prevent this imbalance from affecting the robustness of the network model, it is necessary to collect damaged blade images identified during multiple acquisition processes and then mix them with a suitable proportion of intact blade images to form a dataset. The data in this dataset needs to be labeled for subsequent machine learning use. The standard method and principle of labeling belong to step S2: constructing a convolutional neural network (CNN). The dataset is input into the CNN for training. The CNN has four layers, each containing several modules composed of octave convolutions and channel attention. The input is separated into high-frequency and low-frequency groups after entering the first layer of the network, and then merged in the fourth layer. The merged features are then fed into a fully connected layer to output the classification result. During training, any module first uses octave convolution to extract the input features and divides the extracted features into high-frequency and low-frequency groups. Then, channel attention is used to calibrate the features of the high-frequency and low-frequency groups, and the calibrated two sets of features are fed into the next module.

[0070] In the first three layers of the convolutional neural network, channel attention uses a 2D-DCT-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups, and utilizing more frequency domain implicit information to generalize channel attention to the high-frequency and low-frequency groups to calibrate the feature representations of the high-frequency and low-frequency groups. In the last layer of the network, channel attention uses a GAP-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups to calibrate the feature representations of the high-frequency and low-frequency groups.

[0071] Input X includes high-frequency input X H and low-frequency input X L High-frequency input X HThis refers to pixels where image intensity (brightness / grayscale) changes drastically, such as edges (contours), image details, etc.; low-frequency input X L This refers to pixels with smooth changes in image intensity (brightness / grayscale), such as large areas of color. Wind turbine blades often occupy a larger area in the overall image and have fewer objects and variations compared to ordinary natural images, making it possible to analyze the image in different spatial frequency domains. High-frequency information includes the edges of defects, while low-frequency information includes the approximate location of the defects. The features extracted by octave convolution are divided into high-frequency groups and output Y. H and low-frequency group output Y L .

[0072] The kernel of an octave convolution is divided into four parts, which are represented as follows:

[0073] (1) Convolution kernel W from high frequency group to high frequency group H→H ;

[0074] (2) Convolution kernel W from high frequency group to low frequency group H→L ;

[0075] (3) Convolution kernel W from low-frequency group to high-frequency group L→H ;

[0076] (4) Convolution kernel W from low-frequency group to low-frequency group L→L .

[0077] The convolutional neural network provided in this application consists of four layers. Layers 1, 2, and 3 introduce a channel attention mechanism based on 2D-DCT; layer 4 introduces a channel attention mechanism based on GAP. Within the same layer, high-frequency input X... H The resolution is low frequency input X L 4 times.

[0078] The process of an octave convolutional neural network extracting features from an input image X in a dataset includes:

[0079] High-frequency feature extraction: High-frequency group input X H After convolution kernel W H→H The output Y is obtained after convolution. H→H Low-frequency group input X L First, it passes through the convolution kernel W L→H The convolution operation is followed by upsampling to expand the resolution to the same level as the high-frequency resolution, resulting in the output Y. L→H ;Y H→H and Y L→H After the addition operation, the high-frequency characteristic Y is output. H ;

[0080] Low-frequency feature extraction: low-frequency input X L After convolution kernel WL→L The output Y is obtained after convolution. L→L High-frequency input X H First, the resolution is reduced to the same level as the low-frequency resolution through downsampling, and then it is processed by convolution kernel W. H→L The convolution operation yields the output Y. H →L ;Y L→L With Y H→L After the addition operation, the low-frequency characteristic Y is output. L .

[0081] The calculation processes for high-frequency feature extraction and low-frequency feature extraction are respectively represented as follows:

[0082] Y H =f(X) H W H→H )+upsample(f(X L W L→H ),2);

[0083] Y L =f(X) L W L→L )+f(pool(X H ,2);W H→L );

[0084] In the formula, upsample(·) represents upsampling; pool(·) represents downsampling; and f(·) represents feature extraction.

[0085] While octave convolution enables multi-frequency feature learning, it doesn't focus on specific information. Therefore, attention mechanisms need to be introduced separately in the high-frequency and low-frequency feature learning processes to obtain more discriminative features. Channel attention mechanisms have shown great potential in improving model capabilities. This mechanism can improve feature representation by explicitly constructing various important weights across the channel dimensions. However, channel attention mechanisms are based on GAP (Global Average Pooling), always using a single global average pooling (GAP) to calculate the initial weights for each channel, which may not incorporate other frequency components into the attention process. To address this, we use 2D-DCT (Two-Dimensional Discrete Cosine Transform) to generalize channel attention to high-frequency and low-frequency features with multispectral weights, utilizing more implicit frequency domain information and making it more adaptable to feature groups of different frequencies.

[0086] The transformation formula for 2D-DCT is as follows:

[0087]

[0088] s.tμ∈{0,1,…,A-1},v∈{0,1,…B-1};

[0089] in This represents the input to 2D-DCT. This represents the output of 2D-DCT, where A and B represent the height and width of the 2D-DCT input, respectively, and u and v represent the two-dimensional frequency components, respectively; i and j represent the pixel index, i.e., the pixel in the i-th row and j-th column.

[0090] Use weights To represent the basis functions of 2D-DCT:

[0091]

[0092] The calculation process of 2D-DCT is as follows:

[0093] N frequency components are selected to generalize multi-spectral weights to attract channel attention. The 2D-DCT weights corresponding to the N / 2 higher frequencies among the N frequencies are assigned to the high-frequency group, and the 2D-DCT weights corresponding to the N / 2 lower frequencies among the N frequencies are assigned to the low-frequency group. Each channel in the high-frequency group and the low-frequency group is calculated using the assigned N / 2 2D-DCT weights to obtain one-dimensional feature representations of N / 2 different frequencies.

[0094] Afterwards, the high-frequency group and the low-frequency group each pass through a fully connected layer, which linearly combines their N / 2 one-dimensional feature representations to serve as the final one-dimensional feature representation of the group and passes it to the subsequent channel attention mechanism.

[0095] Regarding the selection of N frequency components, since the feature map output by the last layer is 7×7, it only contains 7×7=49 frequency components. To simplify and improve efficiency, we choose components whose frequency gradually changes in one direction. For example... Figure 5 As shown, the first row represents frequencies that increase in the x-direction but remain unchanged in the y-direction, and the first column represents frequencies that increase in the y-direction but remain unchanged in the x-direction. When testing with N=1, 4, 8, and 16, it was found that the network model performed best when N=16.

[0096] One-dimensional features of the N / 2 frequency components of the l-th channel of the high-frequency group Represented as:

[0097]

[0098] stk∈{1,2,…N / 2}, l∈{1,2,…(1-α)C};

[0099] In the formula, α represents the proportion of the total number of low-frequency channels to the total number of channels. Based on the application scenario, α = 0.5 is selected in this implementation; l represents the channel number, and k represents the frequency number. The 2D-DCT weight of the pixel in the i-th row and j-th column within the k-th frequency component of the high-frequency group is represented by C; C represents the total number of channels.

[0100] One-dimensional feature corresponding to the k-th frequency component of all channels in the high-frequency group Represented as:

[0101]

[0102] In the formula, cat(·) represents a chain operation;

[0103] Freq, the final one-dimensional feature of the entire high-frequency group H Represented as:

[0104]

[0105] In the formula, Represents the fully connected layer coefficients of the high-frequency group;

[0106] One-dimensional feature of the N / 2 frequency components of the l-th channel of the low-frequency group It is expressed as follows:

[0107]

[0108] stk∈{1,2,…N / 2}, l∈{1,2,…αC};

[0109] One-dimensional feature corresponding to the k-th frequency component of all channels in the low-frequency group Represented as:

[0110]

[0111] The final one-dimensional feature Freq of the low-frequency group's full channels L It is expressed as follows:

[0112]

[0113] In the formula, Represents the fully connected layer coefficients of the high-frequency group;

[0114] All channels in the high-frequency and low-frequency groups share the same set of fully connected layer coefficients, which can reduce the number of parameters introduced, decrease the computational resource consumption of the network model, and speed up the network's computation.

[0115] Therefore, the one-dimensional feature of the entire channel is represented as:

[0116] Freq = cat([Freq H ,Freq L ]).

[0117] Freq is a one-dimensional vector of length C, and no channels are reduced during this process, meaning Freq∈R. C×1×1 Next, we use a one-dimensional convolutional kernel of size 3 to capture local cross-channel interactions. Finally, after activating the sigmoid function, we obtain the weights for each channel. The specific calculation of channel attention is as follows:

[0118]

[0119] As can be seen from the above formula, the channel weight ω only considers the interaction between the three neighbors. By appropriately capturing local cross-channel interactions, the computational resource consumption of the model is reduced and the accuracy is increased.

[0120] GAP is a special case of 2D-DCT when u=0, v=0, and its transformation formula is:

[0121]

[0122] The calculation process for GAP is as follows:

[0123]

[0124] Therefore, the one-dimensional feature of the entire channel is represented as:

[0125] Freq=cat([Freq1,Freq2,…Freq C ]);

[0126] Then, we use a one-dimensional convolution kernel of size 3 to capture local cross-channel interactions, and finally, after the sigmoid activation function, we obtain the weights of each channel.

[0127] Step S3: For any wind turbine blade damage detection, input the collected wind turbine blade data into the trained convolutional neural network, and output the judgment result of wind turbine blade damage.

[0128] Data acquisition for wind turbine blades is achieved through an image acquisition system, primarily composed of a drone, a zoom camera, an onboard computer, and control equipment. The drone carries a zoom camera and an onboard computer. The zoom camera captures high-resolution images of the wind turbine blades, while the onboard computer stores these images. Ground control equipment and the drone communicate wirelessly, with the control equipment sending speed, attitude, and position commands to the drone. All modules coordinate to automate the image acquisition process for the wind turbine blades. To obtain high-quality defect data, the wind turbine blades are divided into four regions: the leading edge (windward side), the leeward side, the trailing edge (shaded side), and the leeward side, allowing for multi-angle and multi-directional image capture. During image acquisition, the drone's flight path and shooting points are planned during the initial flight to ensure coverage of all areas of the blades. Afterward, inspection personnel launch the drone, match it to the wind turbine number being inspected, and ensure the drone is fault-free before activating the automatic inspection button. The drone will then take images along the planned flight path. Finally, the drone returns to its designated location, the inspection personnel acquire the captured images, and the drone is shut down.

[0129] This application utilizes octave convolution to allow for inter- and intra-frequency interactions between high-frequency and low-frequency features during feature extraction. It selectively emphasizes useful features and suppresses less useful features in the high-frequency and low-frequency feature portions, respectively. Simultaneously, the low-frequency feature group stores and processes spatially changing features with lower spatial resolution, thereby reducing memory and computational costs and expanding the practical receptive field for defect detection.

[0130] This application also provides a wind turbine blade damage detection system based on a multi-frequency attention network, including:

[0131] The data acquisition module is used to collect historical data of wind turbine blades to form a dataset;

[0132] The network construction module is used to build a convolutional neural network. The dataset is input into the convolutional neural network for training. The convolutional neural network has four layers. Each layer contains several modules composed of octave convolution and channel attention. The input is separated into high-frequency and low-frequency groups after entering the first layer of the network, and then merged in the fourth layer. The merged features are then fed into a fully connected layer to output the classification result. During training, any module first uses octave convolution to extract the input features and divides the extracted features into high-frequency and low-frequency groups. Then, channel attention is used to calibrate the features of the high-frequency and low-frequency groups. The calibrated two sets of features are then fed into the next module as input.

[0133] In the first three layers of the convolutional neural network, channel attention uses a 2D-DCT-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups. This utilizes more frequency domain implicit information to generalize channel attention to the high-frequency and low-frequency groups, thereby calibrating the feature representations of the high-frequency and low-frequency groups. In the last layer, channel attention uses a GAP-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups, thereby calibrating the feature representations of the high-frequency and low-frequency groups.

[0134] The damage discrimination module, for any wind turbine blade damage detection, inputs the collected wind turbine blade data into the trained convolutional neural network and outputs the discrimination result of wind turbine blade damage.

[0135] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for detecting wind turbine blade damage based on a multi-frequency attention network, characterized in that, Includes the following steps: Step S1: Collect historical data of wind turbine blades to form a dataset; Step S2: Construct a convolutional neural network. Input the dataset into the convolutional neural network for training. The convolutional neural network has four layers. Each layer contains several modules composed of octave convolution and channel attention. After being input into the first layer of the network, the data is separated into high-frequency groups and low-frequency groups, and then merged in the fourth layer. The merged features are then fed into a fully connected layer to output the classification result. During training, each module first uses octave convolution to extract input features and divides the extracted features into high-frequency and low-frequency groups. Then, channel attention is used to calibrate the features of the high-frequency and low-frequency groups, and the calibrated two groups of features are fed into the next module as input. In the first three layers of the convolutional neural network, channel attention uses a 2D-DCT-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups. This utilizes more frequency domain implicit information to generalize channel attention to the high-frequency and low-frequency groups, thereby calibrating the feature representations of the high-frequency and low-frequency groups. In the last layer, channel attention uses a GAP-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups, thereby calibrating the feature representations of the high-frequency and low-frequency groups. The calculation process of 2D-DCT is as follows: N frequency components are selected to generalize channel attention. The 2D-DCT weights corresponding to the N / 2 higher frequencies among the N frequencies are assigned to the high-frequency group, and the 2D-DCT weights corresponding to the N / 2 lower frequencies among the N frequencies are assigned to the low-frequency group. Each channel in the high-frequency group and the low-frequency group is calculated using the assigned N / 2 2D-DCT weights to obtain one-dimensional feature representations of N / 2 different frequencies. Step S3: For any wind turbine blade damage detection, input the collected wind turbine blade data into the trained convolutional neural network, and output the judgment result of wind turbine blade damage.

2. The wind turbine blade damage detection method according to claim 1, characterized in that, The kernel of an octave convolution is divided into four parts, which are represented as follows: (1) Convolution kernel from high frequency group to high frequency group ; (2) Convolution kernel from high frequency group to low frequency group ; (3) Convolution kernel from low-frequency group to high-frequency group ; (4) Convolution kernel from low frequency group to low frequency group ; enter Including high frequency input and low frequency input Octave convolutional neural networks process input images in a dataset The feature extraction process includes: High-frequency feature extraction: high-frequency group input After convolution kernel The output is obtained after convolution. Low-frequency group input First, it goes through the convolution kernel. The convolution operation is followed by upsampling to expand the resolution to the same level as the high-frequency group output resolution, resulting in the output. ;Will and The high-frequency characteristics are output after the addition operation. ; Low-frequency feature extraction: low-frequency input After convolution kernel The output is obtained after convolution. High-frequency input First, the resolution is reduced to the same level as the low-frequency input resolution through downsampling, and then it is processed by convolution kernel. Convolution operation to obtain output ;Will and The low-frequency characteristics are output after the addition operation. .

3. The wind turbine blade damage detection method according to claim 2, characterized in that, The transformation formula for 2D-DCT is as follows: in This represents the input to 2D-DCT. This represents the output of 2D-DCT. , These represent the height and width of the 2D-DCT input, respectively. These represent two-dimensional frequency components; Represents the pixel index, i.e., the first pixel. Line 1 The number of pixels in a column; Use weights To represent the basis functions of 2D-DCT: 。 4. The wind turbine blade damage detection method according to claim 3, characterized in that, The first high-frequency group One-dimensional feature of N / 2 frequency components of each channel Represented as: In the formula, This indicates the proportion of low-frequency channels to the total number of channels; Indicates the channel number. Indicates the frequency number; Indicates the high frequency group number Within the frequency component, the first Line 1 2D-DCT weights of column pixels; Indicates the total number of channels; The first of all channels in the high-frequency group One-dimensional feature corresponding to each frequency component Represented as: In the formula, Indicates a serial operation; The final one-dimensional feature of the entire high-frequency group Represented as: In the formula, , ··· Represents the fully connected layer coefficients of the high-frequency group; The first of the low-frequency group One-dimensional features of the N / 2 frequency components of the channel It is expressed as follows: The first of all channels in the low-frequency group One-dimensional feature corresponding to each frequency component Represented as: The final one-dimensional feature of the entire low-frequency group It is expressed as follows: In the formula, , ·· Represents the fully connected layer coefficients of the high-frequency group; Therefore, the one-dimensional feature of the entire channel is represented as: 。 5. The wind turbine blade damage detection method according to claim 4, characterized in that, The calculation process for GAP is as follows: Therefore, the one-dimensional feature of the entire channel is represented as:

6. A wind turbine blade damage detection system based on a multi-frequency attention network, characterized in that, include: The data acquisition module is used to collect historical data of wind turbine blades to form a dataset; The network construction module is used to build a convolutional neural network. The dataset is input into the convolutional neural network for training. The convolutional neural network has four layers. Each layer contains several modules composed of octave convolution and channel attention. The input is separated into high-frequency group and low-frequency group after entering the first layer of the network, and then merged in the fourth layer. The merged features are sent to the fully connected layer to output the classification result. During training, each module first uses octave convolution to extract input features and divides the extracted features into high-frequency and low-frequency groups. Then, channel attention is used to calibrate the features of the high-frequency and low-frequency groups, and the calibrated two groups of features are fed into the next module as input. In the first three layers of the convolutional neural network, channel attention uses a 2D-DCT-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups. This utilizes more frequency domain implicit information to generalize channel attention to the high-frequency and low-frequency groups, thereby calibrating the feature representations of the high-frequency and low-frequency groups. In the last layer, channel attention uses a GAP-based attention mechanism for feature calibration, assigning different weights to each channel of the high-frequency and low-frequency groups, thereby calibrating the feature representations of the high-frequency and low-frequency groups. The calculation process of 2D-DCT is as follows: N frequency components are selected to generalize channel attention. The 2D-DCT weights corresponding to the N / 2 higher frequencies among the N frequencies are assigned to the high-frequency group, and the 2D-DCT weights corresponding to the N / 2 lower frequencies among the N frequencies are assigned to the low-frequency group. Each channel in the high-frequency group and the low-frequency group is calculated using the assigned N / 2 2D-DCT weights to obtain one-dimensional feature representations of N / 2 different frequencies. The damage discrimination module, for any wind turbine blade damage detection, inputs the collected wind turbine blade data into the trained convolutional neural network and outputs the discrimination result of wind turbine blade damage.

Images