Fan gearbox fault diagnosis method, device and system and storage medium

Abstract

The invention discloses a fan gearbox fault diagnosis method, device and system, and a storage medium. The method comprises the following steps: S1, collecting vibration signal data of a gearbox in different fault types; s2, training a double-flow deep neural network based on SG-Net according to the vibration signal data to obtain a fan gearbox fault diagnosis model; the double-flow deep neural network based on the SG-Net comprises a global feature extraction flow network, a local feature extraction flow network and a double-weight adaptive fusion module. And S3, inputting vibration signal data of the fan gear box acquired in real time into the fan gear box fault diagnosis model to obtain a specific fault type of the fan gear box. By adopting the technical scheme of the invention, the accuracy and robustness of fault identification of the fan gearbox are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of fan fault diagnosis, and particularly relates to a fan gearbox fault diagnosis method, device, system, and storage medium. Background Art

[0002] In modern industrial production and energy development, wind power, as a clean and renewable energy source, plays an important strategic supporting role in the transformation of China's energy structure. Due to long-term operation in a complex environment, wind turbines are affected by factors such as high load, vibration, and temperature, and the performance of components such as gearboxes, bearings, and blades will gradually decline, causing various faults, which in turn lead to the shutdown of wind turbines and economic losses. Analyzing the faulty components, the proportion of fan gearbox faults is the highest. Therefore, it is very important to establish a reliable fault diagnosis model for wind turbines.

[0003] For the analysis of rotating component faults, vibration signals are generally used as the signal source. Traditional fault diagnosis methods mainly include time-domain analysis, frequency-domain analysis, and other signal processing-based technologies. Although traditional fault diagnosis methods can identify faults to a certain extent, due to the fact that the vibration signals of fan gearboxes in actual working conditions, especially complex working conditions, are often affected by noise interference, it is difficult to extract and identify fault features, and their diagnostic accuracy and robustness are often limited. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a fan gearbox fault diagnosis method, device, system, and storage medium, which can improve the accuracy and robustness of fan gearbox fault identification, thereby providing a more effective solution for fan gearbox fault diagnosis applications.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A fan gearbox fault diagnosis method includes:

[0007] Step S1: Collect vibration signal data of the gearbox under different fault types;

[0008] Step S2: Train a two-stream deep neural network based on SG-Net according to the vibration signal data to obtain a fan gearbox fault diagnosis model; the two-stream deep neural network based on SG-Net includes: a global feature extraction stream network, a local feature extraction stream network, and a dual-weight adaptive fusion module;

[0009] Step S3: Input the vibration signal data of the fan gearbox collected in real time into the fan gearbox fault diagnosis model to obtain the specific fault type of the fan gearbox.

[0010] Preferably, in step S2, the vibration signal data is processed by a sliding window and transformed by the MTF algorithm to generate a two-dimensional feature map of the vibration signal, and a two-stream deep neural network based on SG-Net is trained according to the two-dimensional feature map of the vibration signal.

[0011] Preferably, the global feature extraction stream network consists of 2 cascaded simplified Swin Transformer encoders. The simplified Swin Transformer encoder includes a Patch Partition module and two stages. Each stage contains two Swin Transformer Blocks. The Swin Transformer Block consists of a normalization layer, a window multi-head self-attention structure, a multi-layer perceptron, a shifted window multi-head self-attention, a normalization layer, and a multi-layer perceptron. The local feature extraction stream network contains two convolutional blocks. Each convolutional block consists of 1 convolutional layer, 1 ReLU activation layer, and 1 max pooling layer. An FCCA module is added after the ReLU activation layer of each convolutional block, and at the same time, the 3*3 convolutional kernel of the convolutional layer is changed to a 5*5 convolutional kernel. The dual-weight adaptive fusion module consists of Cat, a pooling layer, an MLP, and softmax.

[0012] The present invention also provides a fault diagnosis device for a fan gearbox, including:

[0013] An acquisition module for collecting vibration signal data of the gearbox under different fault types;

[0014] A training module for training a two-stream deep neural network based on SG-Net according to the vibration signal data to obtain a fault diagnosis model for the fan gearbox. The two-stream deep neural network based on SG-Net includes: a global feature extraction stream network, a local feature extraction stream network, and a dual-weight adaptive fusion module;

[0015] A diagnosis module for inputting the vibration signal data of the fan gearbox collected in real time into the fault diagnosis model of the fan gearbox to obtain the specific fault type of the fan gearbox.

[0016] Preferably, the training module processes the vibration signal data through a sliding window and transforms it by the MTF algorithm to generate a two-dimensional feature map of the vibration signal, and trains a two-stream deep neural network based on SG-Net according to the two-dimensional feature map of the vibration signal.

[0017] Preferably, the global feature extraction flow network consists of two simplified Swin Transformer encoders connected in series. The simplified Swin Transformer encoder includes a Patch Partition module and two stages. Each stage contains two Swin Transformer Blocks, and each Swin Transformer Block consists of a normalization layer, a window multi-head self-attention structure, a multi-layer perceptron, a shifted window multi-head self-attention, a normalization layer, and a multi-layer perceptron. The local feature extraction flow network contains two convolutional blocks, and each convolutional block consists of one convolutional layer, one ReLU activation layer, and one max pooling layer. An FCCA module is added after the ReLU activation layer of each convolutional block, and at the same time, the 3*3 convolutional kernel of the convolutional layer is changed to a 5*5 convolutional kernel. The dual-weight adaptive fusion module consists of Cat, a pooling layer, an MLP, and softmax.

[0018] An embodiment of the present invention also provides a fan gearbox fault diagnosis system, including: a memory and a processor. A computer program is stored on the memory and run by the processor. When the computer program is run by the processor, it executes the fan gearbox fault diagnosis method.

[0019] An embodiment of the present invention also provides a storage medium. A computer program is stored on the storage medium, and when the computer program runs, it executes the fan gearbox fault diagnosis method.

[0020] The present invention adopts a dual-stream feature extraction network based on Swin transformer and VGGNet16, including a global feature extraction flow and a local feature extraction flow that work in parallel. Global features are extracted by Swin transformer, and local features are extracted by VGGNet16. Secondly, an adaptive weighted fusion module is used to fuse the features extracted by the two feature extraction flows to reduce information loss and provide rich feature information for fault classification. Through advanced feature extraction and fusion technologies, the present invention improves the accuracy and robustness of fan gearbox fault recognition, thus providing a more effective solution for fan gearbox fault diagnosis applications. Description of the Drawings

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0022] Figure 1It is the data processing flowchart of the fault diagnosis method for the fan gearbox of the present invention;

[0023] Figure 2 It is the structure diagram of the Swin Transformer Block;

[0024] Figure 3 It is the structure diagram of the coordinate attention (FCCA) module that fuses channel representations of the present invention;

[0025] Figure 4 It is the structure diagram of the dual-weight adaptive fusion (DWAFN) module designed by the present invention;

[0026] Figure 5 It is the accuracy rate and loss curve graph (training set and test set) of the fan gearbox fault diagnosis method provided by the present invention;

[0027] Figure 6 It is the diagnostic result confusion matrix graph (test set) of the fan gearbox fault diagnosis method provided by the present invention. Detailed implementation manners

[0028] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0029] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0030] Embodiment 1:

[0031] As Figure 1 shown, the embodiment of the present invention provides a fan gearbox fault diagnosis method, including:

[0032] Step 1: Use the acceleration sensors installed on the fan gearbox to collect the vibration signals of the gearbox in different fault types during the offline stage and the vibration signals of the gearbox during the online stage respectively. The offline stage is used for dataset construction and model training, and the online stage is used for real-time data collection and fault diagnosis;

[0033] Step 2: In the offline stage, collect the vibration signal data of the gearbox under different fault types. Group the sample data using a sliding window method, then normalize the grouped data, and divide it into a training set, a validation set, and a test set according to a certain ratio to construct a dataset. Transform the one-dimensional fault signal data in the dataset into two-dimensional feature maps through the Markov transition field (MTF) algorithm and data augmentation technology;

[0034] Step 3: Use a two-stream deep neural network based on SG-Net to perform global feature extraction and local feature extraction on the two-dimensional fault feature data obtained in Step 2, and design a dual-weight adaptive weighted fusion module (DWAFN) to fuse the obtained global features and local features. Finally, combine the global average pooling layer and the fully connected layer to realize the overall design of the fan gearbox fault diagnosis network. Among them, the two-stream deep neural network based on SG-Net includes a global feature extraction stream network, a local feature extraction stream network, and a dual-weight adaptive fusion module (DWAFN);

[0035] Step 4: In the offline stage, input the training set data of the two-dimensional feature maps transformed in Step 2 into the fan gearbox fault diagnosis network proposed in Step 3 for training until the network converges to obtain a fan gearbox fault diagnosis model. Input the test set data into the fan gearbox fault diagnosis model to evaluate the model performance and prevent model overfitting;

[0036] Step 5: In the online stage, collect the vibration signal of the fan gearbox, divide and transform the data in the same way as in Step 2 to obtain two-dimensional feature maps, and then input the two-dimensional feature maps into the fan gearbox fault diagnosis model trained in Step 4 to obtain the online diagnosis result of the fan gearbox fault signal.

[0037] As an implementation manner of the embodiment of the present invention, in Step 2, the sample division of the one-dimensional vibration signals of the fan gearbox under different fault types collected in the offline stage specifically includes:

[0038] S211, Use a sliding window method to divide the dataset into equal-length sequences with a time step of 1024. To prevent the loss of edge information after division, overlap two consecutive signal sequences by 50%.

[0039] S212, Label each fault type and randomly shuffle the sample point order, normalize the sample data, and divide it into a training set, a validation set, and a test set according to a ratio of 7:2:1.

[0040] S213, In S222, use the maximum-minimum normalization method to normalize the sample data, and its formula is:

[0041]

[0042] Among them, x i (k) is the original value of the k-th sample of feature i, and X i，min , x i，max are respectively the minimum value and the maximum value in the samples of feature i, and x i ′ (k) is the normalized value.

[0043] As an implementation manner of an embodiment of the present invention, in step two, converting the fault signal data divided in S211 into a two-dimensional feature map by using the MTF algorithm and data augmentation technology specifically includes:

[0044] S221, generating a two-dimensional feature map with a size of 224×224 by using the MTF algorithm, and converting the generated image from a 4-channel format to a 3-channel RGB format for subsequent image processing and analysis.

[0045] S222, adjusting the resolution of the RGB image to make it more suitable for model training.

[0046] S223, expanding the data set by using scale transformation and random occlusion data augmentation technology to ensure the accuracy and robustness of the fault diagnosis model.

[0047] As an implementation manner of an embodiment of the present invention, in step three, the global feature extraction flow network is constructed based on the Swin transformer encoder, and the specific network structure and global feature extraction include:

[0048] S311. The global feature extraction flow network is used to extract global features and consists of two cascaded simplified SwinTransformer encoders. The Swin transformer encoder network includes an image segmentation block, a stacked module, a normalization layer, a global pooling layer, and a fully connected layer. To avoid the overfitting problem caused by continuous downsampling, the traditional Swintransformer encoder network is simplified to form the global feature extraction flow network, which includes a Patch Partition module and two stages. Assume that the dimension of the two-dimensional feature map input to the network is H×W×C, where H is the height of the feature map, W is the width of the feature map, and C is the number of channels. It is sliced through the Patch Partition module to obtain multiple sub-feature maps, and then the obtained sub-feature maps are input into stage1 in parallel. Stage1 is sequentially cascaded by 1 Linear Embeding module and 2 SwinTransformer Blocks, and Stage2 is sequentially cascaded by 1 Patch Merging module and 2 SwinTransformer Blocks. The Swin Transformer Block consists of a window self-attention module W-MSA, a sliding window self-attention mechanism SW-MSA, a normalization layer LN, and a multi-layer perceptron MLP, as Figure 2 shown, and the Dropout regularization technique is introduced after the W-MSA in Stage1 and Stage2 respectively to improve the generalization ability of the network.

[0049] Furthermore, the window multi-head attention module obtains the attention weights of the model using the scaled dot product operation. The formula for the scaled dot product operation is:

[0050]

[0051] where, is the query matrix, is the key matrix, is the value matrix, represents the scaling factor, m and n are the lengths of the query and the key (or value), and d k and d v represent the dimensions of the query (or key) and the value.

[0052] In S312, the two-dimensional feature map described in Step Four or Step Five is sliced and partitioned using the Patch Partition module to obtain a number of Patches. Each Patch is flattened into a vector by Linear Embeding and mapped to a high-dimensional space. After being processed by LN, MSA, Dropout, LN, and MLP in the first Swin Transformer Block described in S311, it is downsampled by Patch Merging and then processed by the second Swin Transformer Block described in S311 to obtain the global features of the gearbox vibration signal.

[0053] As an implementation manner of the embodiment of the present invention, in Step Three, the local feature extraction flow network is constructed based on VGGNet16, and the specific network structure and local feature extraction include:

[0054] S321, improve the traditional VGGNet16 to form a local feature extraction flow network. First, simplify VGGNet16 into two convolutional blocks, each convolutional block consists of 1 convolutional layer, 1 ReLU activation layer, and 1 max pooling layer. Secondly, improve the coordinate attention mechanism (CA), and propose a fused channel coordinate attention (FCCA) module. Add the FCCA module after the ReLU activation layer of each convolutional block, and change the 3*3 convolutional kernel of the convolutional layer to a 5*5 convolutional kernel.

[0055] S322, process the two-dimensional feature map described in Step Four or Step Five through a convolutional layer, ReLU activation function, FCCA, max pooling layer, etc. to obtain the local features of the gearbox vibration signal.

[0056] Furthermore, the coordinate attention (FCCA) module consists of a three-dimensional information embedding and a three-dimensional attention generation module, as Figure 3 shown, where

[0057] Three-dimensional information embedding module: For each channel of the feature tensor X = [x 1 , x 2 , …, x C ∈ R C×H×W , perform global pooling on the H×W direction using a pooling kernel of size (1,1) for each channel to obtain a channel descriptor, which is expressed by the formula:

[0058]

[0059] Perform global pooling on each channel of the input tensor using pooling kernels of sizes (H,1) and (1,W) in the horizontal and vertical directions to obtain a coordinate descriptor. The output of the c-th channel at height h is:

[0060]

[0061] Similarly, the output of the c-th channel at width w is:

[0062]

[0063] By performing the above pooling operations for feature aggregation in the H, W, and C directions, an attention feature map with direction awareness can be obtained. This coordinate transformation enables the attention module to generate long-range dependencies in the spatial and channel directions, extract precise position information in the spatial and channel directions, and distinguish coordinates with position information.

[0064] 3D attention generation module: The global receptive field of the input features and the encoded precise position information can be obtained by using the aggregated feature map generated by the 3D information embedding module. First, the three generated feature aggregation maps are concatenated, and then the following transformation is performed on the concatenated output:

[0065]

[0066] where F 1 is a 1×1 convolution transformation function, [·, ·] represents the concatenation operation in the spatial dimension, δ is a non-linear activation function, f ∈ R (C / r)×(H+W+1) is the intermediate feature map of the spatial information in the horizontal, vertical, and channel directions, and r is the reduction ratio.

[0067] After that, f is split into three groups of feature maps f h ∈ R (C / r)×h , f w ∈ R (C / r)×w and f c ∈ R (C / r)×1 in the H, W, and C directions. Then, the three groups of feature maps are respectively transformed to the same number of channels as the input X through 1×1 convolutions F h , F w and F c to obtain the attention weights in the H, W, and C directions:

[0068] g h = σ(F h (f h ))

[0069] g w = σ(F w (f w ))

[0070] g c = σ(F c (f c ))

[0071] Among them, σ is the Sigmoid function.

[0072] Finally, multiply the input features by the weights in three directions to obtain the output of the FCCA module, that is:

[0073] y(i,j,k) = x(i,j,k)g h (i)g w (j)g c (k)

[0074] As an implementation manner of the embodiment of the present invention, in step three, the dual-weight adaptive fusion module (DWAFN) is composed of Cat, a pooling layer, MLP, and softmax, as Figure 4 shown. The designed DWAFN specifically includes:

[0075] S331, splice the global features and local features extracted in step three in the channel dimension to obtain spliced features, use adaptive average pooling operation to compress the spliced features, and use the View operation to converge the compressed features into a global feature representation, as shown in the following formula:

[0076] f 12 = View(Pool(Cat(f 1 ,f 2 )))

[0077] Among them, f 1 and w 2 are the global feature and local feature respectively.

[0078] S332, map the global feature obtained by the convergence in S331 through the multi-layer perceptron MLP to obtain the fusion weight. MLP is composed of two fully connected layers FC1 and FC2. In order to increase the non-linear expression ability of MLP, the ReLU activation function is used to connect between the two fully connected layers, that is, the fusion weight is:

[0079] w = FC 2 (ReLU(FC 1 (f 12 )))

[0080] S333, convert the fusion weight obtained in S332 into a probability distribution of the fusion weight through the softmax function to obtain complementary weights, and multiply them by the corresponding global feature and local feature respectively to obtain the final fusion feature, as shown in the following formula:

[0081]

[0082] Among them, F is the fusion feature, represents the tensor multiplication operation.

[0083] As an implementation manner of the embodiment of the present invention, in step four, the fault diagnosis network of the fan gearbox is trained. The learning rate is set to 0.0001, the batch size is 32, and the number of training epochs is 50. The cross-entropy loss is applied to the training of the network until the network converges, and the fault diagnosis model of the fan gearbox is obtained.

[0084] To verify the effectiveness of the method provided by the present invention under the condition of small samples, next, the method of the present invention is proved by using the planetary gearbox data of Southeast University, and an ablation experiment is carried out on this basis. The experimental framework is pytorch-python.

[0085] Application example:

[0086] The vibration data in the fault dataset of the Transmission System Dynamic Simulator (DDS) of Southeast University were collected under two different working conditions: a rotational speed of 1200 r / min and a load of 0 Nm, and a rotational speed of 1800 r / min and a load of 7.32 Nm. Five fault states, including healthy state, broken teeth, missing teeth, tooth root cracks, and tooth surface wear, were collected under each working condition. Each fault type includes 1,048,560 data. The description of the gearbox dataset is shown in Table 1.

[0087] The present invention divides the data in the dataset. The sample size is set to 1165. All samples are converted into two-dimensional feature maps through the MTF algorithm. 815 samples are selected as the training set, 230 samples are selected as the validation set, and 120 samples are selected as the test set. The description of the divided dataset is shown in Table 1.

[0088] Table 1

[0089] Fault type Healthy Defect Broken tooth Tooth root crack Tooth surface wear Label 0 1 2 3 4 Training set 815 815 815 815 815 Validation set 230 230 230 230 230 Test set 120 120 120 120 120

[0090] The diagnosis process of the model can use the preprocessed training set to train the initial network model, and use the validation set for verification to prevent overfitting of the model. Then, the test set is input into the model for testing to verify whether the output result matches the actual fault type. During the training process, the Adam learning rate adaptive algorithm and the cross-entropy loss function are adopted. The number of training epochs is set to 50, the batch size is 64, and the learning rate is 1e-4. The training set is input into the two-stream deep neural network based on SG-Net, and the network is iteratively trained. The accuracy rate and loss rate of each iteration are plotted into a graph, and the result is as Figure 5 shown, and the parameters of the optimal model are saved. The diagnostic accuracy rate of the model is tested using the test set, and the confusion matrix of the diagnostic result is as Figure 6As shown, it can be clearly seen that the recognition rate can reach 100% when identifying health, root cracks, and tooth surface wear. However, when identifying missing teeth and broken teeth faults, 1 sample was misjudged as a health and root crack fault respectively, but the overall model can still achieve a diagnostic accuracy rate of more than 98%, which fully demonstrates that the proposed method has strong feature extraction ability.

[0091] The present invention compares the proposed fault diagnosis model of the fan gearbox based on dual-stream SG-Net and feature fusion with the fault diagnosis models based on the classical deep neural networks Swin Transformer, VGGNet16, Vision Transformer, and ResNet. The test set is used as the input of the diagnosis model, and the diagnostic accuracy rates are shown in Table 2. It can be seen from Table 2 that the accuracy rate of the diagnosis model provided by the present invention is the highest.

[0092] Table 2

[0093]

[0094]

[0095] In order to verify the effectiveness of the FCCA attention module proposed by the present invention, an ablation experiment is carried out. The SG-Net without an attention module (Network A), the SG-Net with a coordinate attention module (Network B), and the SG-Net with an FCCA attention module (Network C) are compared. The test set is used as the input of the diagnosis model, and 5 experiments are carried out respectively. The diagnostic accuracy rates are shown in Table 3. It can be seen from Table 3 that after adopting the FCCA module provided by the present invention, the diagnostic accuracy rate of the model is improved, indicating that the FCCA module provided by the present invention is effective.

[0096] Table 3

[0097]

[0098] Example 2:

[0099] The embodiment of the present invention also provides a fan gearbox fault diagnosis device, including:

[0100] An acquisition module, configured to collect vibration signal data of the gearbox under different fault types;

[0101] A training module, configured to train a dual-stream deep neural network based on SG-Net according to the vibration signal data to obtain a fan gearbox fault diagnosis model; the dual-stream deep neural network based on SG-Net includes: a global feature extraction stream network, a local feature extraction stream network, and a dual-weight adaptive fusion module;

[0102] A diagnostic module, configured to input the vibration signal data of the fan gearbox collected in real time into the fan gearbox fault diagnosis model to obtain the specific fault type of the fan gearbox.

[0103] As an implementation manner of an embodiment of the present invention, the training module generates a two-dimensional feature map of the vibration signal through sliding window processing and MTF algorithm conversion based on the vibration signal data, and trains a two-stream deep neural network based on SG-Net according to the two-dimensional feature map of the vibration signal.

[0104] As an implementation manner of an embodiment of the present invention, the global feature extraction stream network consists of 2 simplified SwinTransformer encoders connected in series. The simplified Swin transformer encoder includes a Patch Partition module and two stages. Each stage contains two Swin Transformer Blocks. The Swin Transformer Block consists of a normalization layer, a window multi-head self-attention structure, a multi-layer perceptron, a shifted window multi-head self-attention, a normalization layer, and a multi-layer perceptron. The local feature extraction stream network contains two convolutional blocks. Each convolutional block consists of 1 convolutional layer, 1 ReLU activation layer, and 1 max pooling layer. An FCCA module is added after the ReLU activation layer of each convolutional block, and at the same time, the 3*3 convolutional kernel of the convolutional layer is changed to a 5*5 convolutional kernel. The dual-weight adaptive fusion module consists of Cat, a pooling layer, an MLP, and a softmax.

[0105] Embodiment 3:

[0106] An embodiment of the present invention further provides a fan gearbox fault diagnosis system, including: a memory and a processor. A computer program is stored on the memory and run by the processor. The computer program, when run by the processor, executes the fan gearbox fault diagnosis method.

[0107] Embodiment 4:

[0108] An embodiment of the present invention further provides a storage medium, on which a computer program is stored. The computer program, when running, executes the fan gearbox fault diagnosis method.

[0109] The above-described embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solution of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for diagnosing a fault in a wind turbine gearbox, characterized in that: include: Step S1, collecting vibration signal data of the gearbox under different fault types; Step S2: training a dual-stream deep neural network based on SG-Net according to the vibration signal data to obtain a fan gearbox fault diagnosis model; the dual-stream deep neural network based on SG-Net includes: a global feature extraction flow network, a local feature extraction flow network and a dual-weight adaptive fusion module; Step S3: input the vibration signal data of the wind turbine gearbox collected in real time into the wind turbine gearbox fault diagnosis model to obtain the specific fault type of the wind turbine gearbox.

2. The wind turbine gearbox fault diagnosis method according to claim 1, characterized in that: In step S2, the vibration signal data is processed by sliding window and converted by MTF algorithm to generate a two-dimensional feature map of the vibration signal, and a two-stream deep neural network based on SG-Net is trained according to the two-dimensional feature map of the vibration signal.

3. The wind turbine gearbox fault diagnosis method according to claim 2, characterized in that: The global feature extraction flow network consists of two simplified Swin Transformer encoders connected in series. The simplified Swin transformer encoder includes a Patch Partition module and two stages. The stage contains two Swin Transformer Blocks. The Swin Transformer Block consists of a standardization layer, a window multi-head self-attention structure, a multi-layer perceptron, a shifted window multi-head self-attention, a normalization layer and a multi-layer perceptron. The local feature extraction flow network contains two convolution blocks, each of which consists of a convolution layer, a ReLU activation layer and a maximum pooling layer. An FCCA module is added after the ReLU activation layer of each convolution block, and the 3*3 convolution kernel of the convolution layer is changed to a 5*5 convolution kernel. The dual-weight adaptive fusion module consists of Cat, a pooling layer, MLP and softmax.

4. A fan gearbox fault diagnosis device, characterized in that: include: An acquisition module is used to collect vibration signal data of the gearbox under different fault types; A training module is used to train a dual-stream deep neural network based on SG-Net according to vibration signal data to obtain a wind turbine gearbox fault diagnosis model; The two-stream deep neural network based on SG-Net includes: global feature extraction stream network, local feature extraction stream network and dual-weight adaptive fusion module; The diagnosis module is used to input the vibration signal data collected in real time from the fan gearbox into the fan gearbox fault diagnosis model to obtain the specific fault type of the fan gearbox.

5. The wind turbine gearbox fault diagnosis device according to claim 4, characterized in that: The training module generates a two-dimensional feature map of the vibration signal through sliding window processing and MTF algorithm conversion according to the vibration signal data, and trains a two-stream deep neural network based on SG-Net according to the two-dimensional feature map of the vibration signal.

6. The wind turbine gearbox fault diagnosis device according to claim 5, characterized in that: The global feature extraction flow network consists of two simplified Swin Transformer encoders connected in series. The simplified Swin transformer encoder includes a Patch Partition module and two stages. The stage contains two Swin Transformer Blocks. The Swin Transformer Block consists of a standardization layer, a window multi-head self-attention structure, a multi-layer perceptron, a shifted window multi-head self-attention, a normalization layer and a multi-layer perceptron. The local feature extraction flow network contains two convolution blocks, each of which consists of a convolution layer, a ReLU activation layer and a maximum pooling layer. An FCCA module is added after the ReLU activation layer of each convolution block, and the 3*3 convolution kernel of the convolution layer is changed to a 5*5 convolution kernel. The dual-weight adaptive fusion module consists of Cat, a pooling layer, MLP and softmax.

7. A fan gearbox fault diagnosis system, characterized in that: include: A memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, the wind turbine gearbox fault diagnosis method according to any one of claims 1 to 3 is executed.

8. A storage medium, characterized in that: The storage medium stores a computer program, and the computer program executes the wind turbine gearbox fault diagnosis method according to any one of claims 1 to 3 when running.

Images