Fault diagnosis method for wind power bearing based on dual-channel feature fusion

Abstract

The application provides a kind of double channel feature fusion wind power bearing fault diagnosis method, comprising the following steps: obtaining bearing vibration data, the bearing vibration data is divided into source domain data and target domain data;According to source domain data and target domain data, based on the BLS model and multi-scale hollow attention structure of optimization extraction global feature;The short-time Fourier transform processing is carried out to source domain data and target domain data, and time-frequency feature map is generated;Time-frequency feature map is input into hollow convolutional neural network, and multi-scale local feature is extracted;The global feature and multi-scale local feature obtained are spliced to obtain the feature vector for fault identification;The feature vector is input into the field self-adaptive module for field self-adaptive adjustment, to minimize the difference between source domain and target domain feature distribution;The feature vector after field self-adaptive adjustment is input into the classifier, and the classifier outputs the probability that the target domain data sample belongs to each fault category.The application can improve the fault identification capability.

Description

Technical Field

[0001] This invention relates to a fault diagnosis method, and more particularly to a fault diagnosis method for wind turbine bearings using dual-channel feature fusion. Background Technology

[0002] As a critical supporting component in wind turbines, the operating status of bearings directly affects the stability and safety of the equipment. However, most existing bearing fault diagnosis methods have limited feature extraction capabilities, and existing diagnostic models mostly adopt fixed structures, lacking the ability to adaptively fuse local features and multi-scale information. This leads to unstable performance under complex operating conditions and makes it difficult to meet the accurate diagnostic needs under small sample and heterogeneous data conditions in complex operating environments. Summary of the Invention

[0003] To address at least one technical problem in the prior art, embodiments of the present invention provide a wind turbine bearing fault diagnosis method based on dual-channel feature fusion, thereby improving fault identification capability and robustness. To achieve the above technical objectives, the technical solution adopted in embodiments of the present invention is as follows:

[0004] This invention provides a method for diagnosing wind turbine bearing faults using dual-channel feature fusion, comprising the following steps:

[0005] Step S10: Obtain bearing vibration data, which is divided into source domain data and target domain data; the source domain data is historical data, and the target domain data is current data.

[0006] Step S20: Extract global features based on the source domain data and target domain data, using an optimized BLS model and a multi-scale dilated attention structure.

[0007] Step S30: Perform short-time Fourier transform on the source domain data and the target domain data to generate a time-frequency feature map; input the time-frequency feature map into a dilated convolutional neural network to extract multi-scale local features;

[0008] Step S40: The obtained global features and multi-scale local features are concatenated to obtain a feature vector for fault identification;

[0009] Step S50: Input the feature vector into the domain adaptation module for domain adaptation adjustment to minimize the difference in feature distribution between the source domain and the target domain, thereby aligning the feature distributions of the source domain and the target domain.

[0010] Step S60: Input the adaptively adjusted feature vector into the classifier, and the classifier outputs the probability that the target domain data sample belongs to each fault category.

[0011] Furthermore, step S20 specifically includes:

[0012] Step S201: Based on the source domain data, optimize the BLS model using the Grey Wolf optimization algorithm; including:

[0013] The gray wolf optimization algorithm guides the entire wolf pack to gradually approach the optimal solution by continuously simulating the hunting strategies and social hierarchy of α, β, and δ wolves. First, the gray wolf pack is initialized, and each gray wolf contains a set of BLS model structure parameters, including the number of feature nodes, the number of augmentation nodes, and the sparse regularization factor λ.

[0014] The source domain data is divided into a training set and a test set;

[0015] For each individual gray wolf, a corresponding BLS model structure is constructed. Given the dataset X in the input training set, the feature nodes Z of the BLS model constructed for each individual gray wolf are... i Represented as: z i =φ(XW zi +β zi ), i = 1, 2, ..., n, where W zi and β zi Let Z represent the weight matrix and bias matrix randomly generated at the i-th feature node of the BLS model, respectively. Let φ represent the linear activation function, and Z represent the bias matrix. n Z represents the feature node vector composed of the generated feature nodes. n ≡[z1…z n Then, feature node z is established. i To enhance node h j The mapping; where the enhancement node h j The output formula is: h j =ξ(Z) n W hj +β hj ), j=1,...,m, where W hj and β hj Let represent the weight matrix and bias matrix randomly generated by the j-th booster node, respectively, and ξ represent the nonlinear activation function, which applies to all generated booster nodes h. j Constructing the enhanced node vector H m H m ≡[h1…h m ], the feature node vector Z n and the enhanced node vector H m Matrix concatenation yields the joint feature matrix A = [Z] for each individual gray wolf. n |H m The joint feature matrix A is used as the input layer of the BLS model to train the BLS model, and the output weight matrix W of the BLS model is calculated and updated.

[0016] In the optimization process of the BLS model, a sparse regularization mechanism is adopted to improve the robustness of the model and suppress redundant features, and the weight matrix W is output by constraining the l1 norm.

[0017] In the optimization process of the BLS model, feature selection is performed by minimizing the following objective function:

[0018]

[0019] Where A and b represent the joint feature matrix of the feature node and the enhancement node and the real fault label matrix, respectively, W represents the output weight matrix that needs to be optimized, and λ is the sparsity regularization factor that controls the sparsity.

[0020] The classification accuracy of the BLS model structure parameters on the validation set is used as the fitness function. The positions of α wolf, β wolf and δ wolf are updated iteratively until the maximum number of iterations is reached or the early termination condition is met. The position of α wolf with the highest fitness ranking is obtained as the output. Finally, the BLS model structure parameters that maximize the fault diagnosis classification accuracy are obtained.

[0021] Step S202: Input the source domain data and target domain data into the optimized BLS model for feature extraction to obtain a sparse global feature matrix;

[0022] Step S203 involves inputting the obtained sparse global feature matrix into a multi-scale dilated attention structure for processing, generating global features that fuse multi-scale information; including:

[0023] First, the multi-scale dilated attention structure has three parallel dilated convolutional layers, each with an inflation rate of 1, 2, and 3, respectively, to capture global features. The output features of the three dilated convolutional layers are concatenated to obtain a feature map. Second, the feature map channel is divided into multiple heads, each of which independently calculates attention weights to focus on fault features in different frequency bands or time periods, and applies different inflation rates to each head. Then, within each head, a local window self-attention operation is performed on the feature map.

[0024] Furthermore, in step S201, the range of the BLS model structure parameters is set as follows: the number of feature nodes is between [50, 200], the number of augmentation nodes is between [100, 500], and the sparse regularization factor λ is between [0.01, 1].

[0025] Furthermore, in step S201, the early termination condition is configured to terminate early when the classification accuracy on the validation set improves by less than 0.1% over 20 consecutive iterations.

[0026] Further, in step S203, the local window self-attention operation includes:

[0027] Each head uses scaled dot product attention, represented as follows:

[0028] h i =SDPA(Q i ,K i V i ,r i ), 1≤i≤n

[0029] Among them, h i This represents the header of the feature mapping channel division, r i Let Q represent the expansion rate applicable to the i-th head, and Q... i K i and V i This represents the feature map segment assigned to it; the output of each head. The data is concatenated and forwarded through a linear layer for global feature fusion.

[0030] Furthermore, step S30 specifically includes:

[0031] Step S301, the short-time Fourier transform processing of the source domain data and the target domain data is as follows:

[0032]

[0033] Where x(t) represents the input data, w(tm) represents the sliding window function centered at m, ω is the angular frequency, and e -jωt Indicates the frequency domain transform factor;

[0034] Step S302, the step of inputting the time-frequency feature map into a dilated convolutional neural network to extract multi-scale local features includes:

[0035] The dilated convolutional neural network comprises three one-dimensional dilated convolutional layers, each with a different dilation rate of 1, 2, and 3. The corresponding convolution operations sample input features from adjacent points, with a gap of 1 point, and a gap of 2 points, respectively. After the three one-dimensional dilated convolutional layers is a fully connected layer. Each one-dimensional dilated convolutional layer retains the same output feature dimension after convolution. Finally, the local features extracted by the three one-dimensional dilated convolutional layers are concatenated and fused, and then fed into the fully connected layer for linear combination to obtain multi-scale local features.

[0036] Furthermore, step S50 specifically includes:

[0037] Step S501: First, align the global feature distributions of the source and target domains.

[0038] Calculate distance L mmd :

[0039]

[0040] Among them, X S 'and X T ' represents the source domain features and the target domain features, respectively, where n1 represents the number of source domain features and n2 represents the number of target domain features. Represents source domain feature samples, Represents the feature samples of the target domain. It is a mapping function that projects both source and target domain features into a high-dimensional Hilbert space, and calculates the distance L in the high-dimensional Hilbert space. mmd Minimize this distance L mmd Achieve global feature distribution alignment between the source and target domains;

[0041] Step S502, next, align the class-level feature distributions of the source domain and the target domain;

[0042] Calculate distance L lmmd :

[0043]

[0044] C represents the number of subdomain categories. and D represents the source domain class label sample and the target domain class label sample, respectively; s D represents the set of source domain feature samples within the subdomain. t This represents the set of feature samples of the target domain within the subdomain;

[0045] Minimize this distance L lmmd Achieve alignment of class-level feature distributions between the source and target domains.

[0046] The beneficial effects of the technical solution provided by the embodiments of the present invention are as follows:

[0047] 1) By optimizing the width learning system model and multi-scale hollow attention structure, the problems of existing methods being limited by single-scale expressive ability and weak ability to identify essential features under complex working conditions during fault feature extraction are solved; at the same time, the multi-scale hollow attention structure can expand the receptive field and capture key feature information at multiple scales.

[0048] 2) Based on the time-frequency feature map obtained by short-time Fourier transform, dilated convolutional neural networks are used to extract multi-scale local features. This can expand the receptive field of the network without increasing the number of parameters, and effectively capture causal information across scales and time periods.

[0049] 3) Through domain adaptation, knowledge from the source domain can be transferred to the target domain, enabling high-precision cross-domain fault diagnosis. Attached Figure Description

[0050] Figure 1 This is a flowchart of the diagnostic method in an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of a dilated convolutional neural network in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] In the description of the embodiments of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0054] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can also refer to the internal connection of two components; and they can refer to a wireless connection or a wired connection. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0055] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0056] like Figure 1 As shown in the figure, this invention proposes a method for fault diagnosis of wind turbine bearings based on dual-channel feature fusion, comprising the following steps:

[0057] Step S10: Obtain bearing vibration data, which is divided into source domain data and target domain data; the source domain data is historical data, and the target domain data is current data.

[0058] Step S20: Extract global features based on the source domain data and target domain data, using an optimized BLS model and a multi-scale dilated attention structure.

[0059] BLS stands for Broad Learning System; MSDA stands for Multi-scale Dilated Attention.

[0060] Step S20 specifically includes:

[0061] Step S201: Based on the source domain data, optimize the BLS model using the Grey Wolf optimization algorithm; including:

[0062] The gray wolf optimization algorithm guides the entire wolf pack to gradually approach the optimal solution by continuously simulating the hunting strategies and social hierarchy of α, β, and δ wolves. First, the gray wolf pack is initialized, and each gray wolf contains a set of BLS model structure parameters, including the number of feature nodes, the number of augmentation nodes, and the sparse regularization factor λ.

[0063] The source domain data is divided into a training set and a test set;

[0064] For each individual gray wolf, a corresponding BLS model structure is constructed. Given the dataset X in the input training set, the feature nodes Z of the BLS model constructed for each individual gray wolf are... i Represented as: z i =φ(XW zi +β zi ), i = 1, 2, ..., n, where W zi and β zi Let Z represent the weight matrix and bias matrix randomly generated at the i-th feature node of the BLS model, respectively. Let φ represent the linear activation function, and Z represent the bias matrix. n Z represents the feature node vector composed of the generated feature nodes. n ≡[z1…z n Then, feature node z is established. i To enhance node h j The mapping; where the enhancement node h j The output formula is: h j =ξ(Z) n W hj +β hj ), j=1,...,m, where W hj and β hj Let represent the weight matrix and bias matrix randomly generated by the j-th booster node, respectively, and ξ represent the nonlinear activation function, which applies to all generated booster nodes h. j Constructing the enhanced node vector H m H m ≡[h1…h m ], the feature node vector Z nand the enhanced node vector H m Matrix concatenation yields the joint feature matrix A = [Z] for each individual gray wolf. n |H m The joint feature matrix A is used as the input layer of the BLS model to train the BLS model, and the output weight matrix W of the BLS model is calculated and updated.

[0065] In the optimization process of the BLS model, a sparse regularization mechanism is adopted to improve the robustness of the model and suppress redundant features. The weight matrix W is output by constraining the l1 norm. This can enhance the feature selection capability and ensure that only the key feature channels that have a significant contribution to the fault diagnosis results are retained.

[0066] In the optimization process of the BLS model, feature selection is performed by minimizing the following objective function:

[0067]

[0068] Where A and b represent the joint feature matrix of the feature node and the enhancement node and the real fault label matrix, respectively, W represents the output weight matrix that needs to be optimized, and λ is the sparsity regularization factor that controls the sparsity.

[0069] The classification accuracy of the BLS model structure parameters on the validation set is used as the fitness function. The positions of α wolf, β wolf and δ wolf are updated iteratively until the maximum number of iterations is reached or the early termination condition is met. The position of α wolf with the highest fitness ranking is obtained as the output. Finally, the BLS model structure parameters that maximize the fault diagnosis classification accuracy are obtained.

[0070] Preferably, the range of BLS model structural parameters is set as follows: the number of feature nodes is between [50, 200], the number of augmentation nodes is between [100, 500], and the sparsity regularization factor λ is between [0.01, 1].

[0071] Preferably, the early termination condition is configured to terminate early when the classification accuracy on the validation set improves by less than 0.1% over 20 consecutive iterations, in order to balance optimization efficiency and diagnostic performance.

[0072] Step S202: Input the source domain data and target domain data into the optimized BLS model for feature extraction to obtain a sparse global feature matrix;

[0073] Step S203 involves inputting the obtained sparse global feature matrix into a multi-scale dilated attention structure for processing, generating global features that fuse multi-scale information; including:

[0074] First, the multi-scale dilated attention structure has three parallel dilated convolutional layers, each with a dilation rate of 1, 2, and 3 to capture global features. The output features of the three dilated convolutional layers are concatenated to obtain a feature map. Second, the feature map channel is divided into multiple heads, each of which independently calculates attention weights to focus on fault features in different frequency bands or time periods, and applies different dilation rates to each head. Then, within each head, a local window self-attention operation is performed on the feature map.

[0075] The local window self-attention operation includes:

[0076] Each head uses Scaled Dot Product Attention (SDPA), represented as follows:

[0077] h i =SDPA(Q i ,K i V i ,r i ), 1≤i≤n

[0078] Among them, h i This represents the header of the feature mapping channel division, r i Let Q represent the expansion rate applicable to the i-th head, and Q... i K i and V i This represents the feature map segment assigned to it; the output of each head. The data is concatenated and forwarded through a linear layer for global feature fusion;

[0079] Step S30: Perform short-time Fourier transform on the source domain data and the target domain data to generate a time-frequency feature map; input the time-frequency feature map into a dilated convolutional neural network to extract multi-scale local features;

[0080] Step S301, the short-time Fourier transform processing of the source domain data and the target domain data is as follows:

[0081]

[0082] Where x(t) represents the input data, w(tm) represents the sliding window function centered at m, ω is the angular frequency, and e -jωt Indicates the frequency domain transform factor;

[0083] The short-time Fourier transform divides the bearing vibration data into frames by setting a fixed window length and overlap rate. A Fourier transform is applied to each frame to generate a two-dimensional time-frequency feature map containing time and frequency dimensions, which is used to characterize the transient feature information of non-stationary fault signals.

[0084] Step S302, the step of inputting the time-frequency feature map into a dilated convolutional neural network to extract multi-scale local features includes:

[0085] like Figure 2 As shown, the dilated convolutional neural network includes three one-dimensional dilated convolutional layers. Each one-dimensional dilated convolutional layer uses a different dilation rate, set to 1, 2, and 3 respectively. The corresponding convolution operations sample input features from adjacent points, with a gap of 1 point, and a gap of 2 points respectively. After the three one-dimensional dilated convolutional layers, there is a fully connected layer. Each one-dimensional dilated convolutional layer retains the same output feature dimension after convolution. Finally, the local features extracted by the three one-dimensional dilated convolutional layers are concatenated and fused, and then fed into the fully connected layer for linear combination to obtain multi-scale local features.

[0086] Step S40: The obtained global features and multi-scale local features are concatenated to obtain a feature vector for fault identification;

[0087] Step S50 involves inputting the feature vector into the domain adaptation module for domain adaptation adjustment to minimize the difference in feature distribution between the source and target domains, thereby aligning the feature distributions of the source and target domains; including:

[0088] Step S501: First, align the global feature distributions of the source and target domains.

[0089] Calculate distance L mmd :

[0090]

[0091] Among them, X S 'and X T ' represents the source domain features and the target domain features, respectively, where n1 represents the number of source domain features and n2 represents the number of target domain features. Represents source domain feature samples, Represents the feature samples of the target domain. It is a mapping function that projects both source and target domain features into a high-dimensional Hilbert space, and calculates the distance L in the high-dimensional Hilbert space. mmd Minimize this distance L mmd Achieve global feature distribution alignment between the source and target domains;

[0092] Step S502, next, align the class-level feature distributions of the source domain and the target domain;

[0093] Calculate distance L lmmd :

[0094]

[0095] C represents the number of subdomain categories. and D represents the source domain class label sample and the target domain class label sample, respectively; s D represents the set of source domain feature samples within the subdomain. t This represents the set of feature samples of the target domain within the subdomain;

[0096] Minimize this distance L lmmd Achieve alignment of class-level feature distributions between the source and target domains;

[0097] Step S60: Input the adaptively adjusted feature vector into the classifier, and the classifier outputs the probability that the target domain data sample belongs to each fault category;

[0098] The classifier can use a fully connected layer + Softmax structure;

[0099] This application achieves high-precision cross-domain fault diagnosis by using dual-channel feature fusion and domain adaptation to transfer knowledge from the source domain to the target domain.

[0100] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for fault diagnosis of wind turbine bearings using dual-channel feature fusion, characterized in that, Includes the following steps: Step S10: Obtain bearing vibration data, which is divided into source domain data and target domain data; the source domain data is historical data, and the target domain data is current data. Step S20: Extract global features based on the source domain data and target domain data, using an optimized BLS model and a multi-scale dilated attention structure. Step S20 specifically includes: Step S201: Optimize the BLS model using the Grey Wolf optimization algorithm based on the source domain data; Step S202: Input the source domain data and target domain data into the optimized BLS model for feature extraction to obtain a sparse global feature matrix; Step S203 involves inputting the obtained sparse global feature matrix into a multi-scale dilated attention structure for processing, generating global features that fuse multi-scale information; including: First, the multi-scale dilated attention structure has three parallel dilated convolutional layers, each with a dilation rate of 1, 2, and 3 to capture global features. The output features of the three dilated convolutional layers are concatenated to obtain a feature map. Second, the feature map channel is divided into multiple heads, each of which independently calculates attention weights to focus on fault features in different frequency bands or time periods, and applies different dilation rates to each head. Then, within each head, a local window self-attention operation is performed on the feature map. In step S203, the local window self-attention operation includes: Each head uses scaled dot product attention, represented as follows: ； in, This represents the header of the feature mapping channel segmentation. Indicates that it applies to the first The expansion rate of the head, and , and This represents the feature map segment assigned to it; the output of each head. The data is concatenated and forwarded through a linear layer for global feature fusion; Step S30: Perform short-time Fourier transform on the source domain data and the target domain data to generate a time-frequency feature map; input the time-frequency feature map into a dilated convolutional neural network to extract multi-scale local features; Step S40: The obtained global features and multi-scale local features are concatenated to obtain a feature vector for fault identification; Step S50: Input the feature vector into the domain adaptation module for domain adaptation adjustment to minimize the difference in feature distribution between the source domain and the target domain, thereby aligning the feature distributions of the source domain and the target domain. Step S60: Input the adaptively adjusted feature vector into the classifier, and the classifier outputs the probability that the target domain data sample belongs to each fault category.

2. The wind turbine bearing fault diagnosis method based on dual-channel feature fusion as described in claim 1, characterized in that, Step S201 involves optimizing the BLS model using the Grey Wolf optimization algorithm based on the source domain data; including: The Grey Wolf optimization algorithm simulates through continuous iteration. Wolf, wolves and The wolf pack's hunting strategies and social hierarchy guide the entire pack to gradually approach the optimal solution. First, the gray wolf pack is initialized, with each individual wolf containing a set of BLS model structural parameters, including the number of feature nodes, the number of augmentation nodes, and the sparsity regularization factor. ; The source domain data is divided into a training set and a test set; For each individual gray wolf, a corresponding BLS model structure is constructed. Given the dataset X in the input training set, the feature nodes of the BLS model constructed for each individual gray wolf are... Represented as: In the formula, and These represent the first and second digits of the BLS model, respectively. The weight matrix and bias matrix are randomly generated for each feature node. Represents a linear activation function. This represents the feature node vector composed of the generated feature nodes. Then, feature nodes are established. To the enhanced node The mapping; where, the enhanced node The output formula is: ,in and respectively represent the first The weight matrix and bias matrix are randomly generated by each enhancement node, and This represents a non-linear activation function that generates all the enhanced nodes. Constructing enhanced node vectors , , feature node vector and enhanced node vectors Perform matrix concatenation to obtain the joint feature matrix corresponding to each individual gray wolf. The joint feature matrix A The BLS model is trained using the input layer, and the output weight matrix of the BLS model is calculated and updated. ; In the optimization process of the BLS model, a sparse regularization mechanism is used to improve the model's robustness and suppress redundant features. Norm constraint output weight matrix ; In the optimization process of the BLS model, feature selection is performed by minimizing the following objective function: ； in, and These represent the joint feature matrix of the feature node and the enhancement node, and the true fault label matrix, respectively. This represents the output weight matrix that needs to be optimized. A sparsity regularization factor to control the degree of sparsity; The classification accuracy of the BLS model's structural parameters on the validation set is used as the fitness function, and the model is updated iteratively. Wolf, wolves and The wolf's position is determined until the maximum number of iterations is reached or the early termination condition is met, at which point the wolf with the highest fitness ranking is obtained. The wolf's position is used as the output, and the BLS model structure parameters that maximize the accuracy of fault diagnosis classification are finally obtained.

3. The wind turbine bearing fault diagnosis method based on dual-channel feature fusion as described in claim 2, characterized in that, In step S201, the range of BLS model structure parameters is set as follows: the number of feature nodes is between [50, 200], the number of augmentation nodes is between [100, 500], and the sparsity regularization factor is... Between [0.01, 1].

4. The wind turbine bearing fault diagnosis method based on dual-channel feature fusion as described in claim 2, characterized in that, In step S201, the early termination condition is configured to terminate early when the classification accuracy on the validation set improves by less than 0.1% over 20 consecutive iterations.

5. The wind turbine bearing fault diagnosis method based on dual-channel feature fusion as described in any one of claims 1 to 4, characterized in that, Step S30 specifically includes: Step S301, the short-time Fourier transform processing of the source domain data and the target domain data is as follows: ； in, This represents the input data. Indicated by The sliding window function centered on the center, Angular frequency, Indicates the frequency domain transform factor; Step S302, the step of inputting the time-frequency feature map into a dilated convolutional neural network to extract multi-scale local features includes: The dilated convolutional neural network comprises three one-dimensional dilated convolutional layers, each with a different dilation rate of 1, 2, and 3. The corresponding convolution operations sample input features from adjacent points, with a gap of 1 point, and a gap of 2 points, respectively. After the three one-dimensional dilated convolutional layers is a fully connected layer. Each one-dimensional dilated convolutional layer retains the same output feature dimension after convolution. Finally, the local features extracted by the three one-dimensional dilated convolutional layers are concatenated and fused, and then fed into the fully connected layer for linear combination to obtain multi-scale local features.

6. The wind turbine bearing fault diagnosis method based on dual-channel feature fusion as described in any one of claims 1 to 4, characterized in that, Step S50 specifically includes: Step S501: First, align the global feature distributions of the source and target domains. Calculate distance : ； in, and Representing source domain features and target domain features respectively, Indicates the number of source domain features. Indicates the number of features in the target domain. Represents source domain feature samples, Represents the feature samples of the target domain. It is a mapping function that projects both source and target domain features into a high-dimensional Hilbert space, and calculates the distance in the high-dimensional Hilbert space. Minimize this distance Achieve global feature distribution alignment between the source and target domains; Step S502, next, align the class-level feature distributions of the source domain and the target domain; Calculate distance ; ； Indicates the number of subdomain categories. and These represent source domain class label samples and target domain class label samples, respectively. This represents the set of source domain feature samples within the subdomain. This represents the set of feature samples of the target domain within the subdomain; Minimize this distance Achieve alignment of class-level feature distributions between the source and target domains.

Images