A fan bearing fault diagnosis method based on double-branch spatio-temporal feature fusion

Abstract

This invention discloses a wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion, belonging to the field of bearing fault diagnosis technology. It addresses the problems of insufficient utilization of temporal information and weak key feature focusing ability in traditional fault diagnosis models. This invention designs a dual-branch structure: one branch uses a convolutional neural network to extract spatial features, supplemented by an improved multi-scale deep separable convolutional block attention module to enhance the focusing ability on key features; the other branch mines long temporal dependency information through a Transformer model, introducing a bidirectional long short-term memory network to compensate for its deficiency in modeling local temporal dynamic characteristics. This solves the problem of insufficient utilization of temporal information in traditional fault diagnosis models. Finally, this invention integrates the spatiotemporal features extracted by the two branches using a bidirectional interactive fusion mechanism, improving the performance and generalization ability of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and in particular to a method for diagnosing wind turbine bearing faults based on dual-branch spatiotemporal feature fusion. Background Technology

[0002] Modern mechanical equipment is becoming increasingly complex in structure and its manufacturing precision is constantly improving. An anomaly in a single component can trigger a systemic chain of failures. As a core supporting component of rotating machinery, rolling bearings face a significantly increased risk of failure when subjected to extreme service conditions such as high loads and high temperatures over extended periods. To ensure the safe operation of equipment, accurate identification of bearing faults is crucial. Technical research in this field mainly revolves around two main lines: first, extracting and reconstructing characteristic information representing faults from raw monitoring data; and second, classifying and identifying fault modes. The former constitutes a prerequisite for bearing fault identification, relying on signal analysis methods to extract sufficient fault characteristics from the basic vibration response, providing a data foundation for subsequent processing.

[0003] Current technological approaches primarily rely on vibration signal analysis, collecting operational status data through acceleration, velocity, or displacement sensors. This data is then combined with signal processing techniques (such as time-frequency analysis, wavelet transform, and empirical mode decomposition) and machine learning algorithms (support vector machines, random forests, and deep learning) to uncover fault features and identify patterns. This process is essentially a closed loop of "information extraction - intelligent decision-making," where data feature extraction, as the first step in the diagnostic chain, directly determines the upper limit of accuracy for subsequent classification and identification. However, this technology still faces severe challenges in practical engineering applications: First, the inherent non-stationarity (time-varying statistical characteristics due to operating condition switching) and multi-component coupling (fault impact, harmonic interference, and environmental noise aliasing) of bearing vibration signals make it difficult for traditional signal processing methods to accurately decouple fault components, easily missing subtle features of early weak faults and resulting in incomplete feature extraction. Second, traditional fault diagnosis models (such as shallow machine learning models) do not make sufficient use of temporal information, treating vibration sequences as isolated samples or sets of statistical features, ignoring the dynamic dependencies and long short-term memory characteristics in the fault evolution process, and making it difficult to capture the temporal patterns of fault development. Third, existing models generally lack the ability to focus on key features, treating the massive amounts of extracted high-dimensional features indiscriminately, failing to adaptively strengthen discriminative features strongly correlated with fault modes, or suppress redundant or interfering features, resulting in insufficient robustness of the model under complex operating conditions and limited diagnostic accuracy and generalization ability. In summary, the incomplete feature extraction caused by the non-stationarity and multi-component coupling of rolling bearing vibration signals, as well as the problems of insufficient utilization of temporal information and weak ability to focus on key features in traditional fault diagnosis models, all contribute to the challenges.

[0004] Therefore, a method for diagnosing wind turbine bearing faults based on the fusion of dual-branch spatiotemporal features is needed. Summary of the Invention

[0005] To address the problems of incomplete feature extraction from rolling bearing vibration signals due to non-stationarity and multi-component coupling, as well as the insufficient utilization of temporal information and weak key feature focusing ability of traditional fault diagnosis models, this invention provides a wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion. By designing a dual-branch structure, one branch utilizes CNN to extract spatial features, supplemented by an improved Multi-scale Depthwise Separable Convolution Block Attention Module (MDSCBAM) to enhance the focusing ability on key features; the other branch mines long-term temporal dependency information using a Transformer model, introducing BiLSTM to compensate for its shortcomings in modeling local temporal dynamic characteristics, thus solving the problem of insufficient utilization of temporal information in traditional fault diagnosis models. At the input data level, the original vibration signal is preprocessed using a combination of VMD and FFT to improve the completeness of feature extraction. Finally, a bidirectional interactive fusion mechanism is used to integrate the spatiotemporal features extracted from the two branches, improving the performance and generalization ability of fault diagnosis. The specific technical solution is as follows: A wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion includes the following steps: S1: Perform preprocessing operations on the input signal, wherein the preprocessing operations include at least variable mode decomposition and fast Fourier transform; S2: Input the multi-scale signal tensor generated by the fusion of variational mode decomposition and fast Fourier transform, and output the probability distribution of fault categories using the fault diagnosis model; The fault diagnosis model includes a spatial feature branch, a temporal feature branch, and a fusion mechanism for the spatial feature branch and the temporal feature branch. The spatial feature branch uses multi-level convolutional blocks for temporal compression and feature mapping. The temporal feature branch includes the encoder part of the Transformer and a bidirectional long short-term memory network. The fusion mechanism is specifically a bidirectional interactive fusion mechanism. The bidirectional interactive fusion mechanism generates query vectors to obtain information from the temporal features, or responds to the query of the temporal features as key values, forming a bidirectional information flow. Finally, the fusion interaction result is combined with the original spatiotemporal features.

[0006] Preferably, the variable mode decomposition achieves signal decomposition by solving a constrained optimization problem, and an augmented Lagrangian function is introduced to transform the constrained problem into an unconstrained optimization problem. The augmented Lagrangian function is as follows: in, This is a regularization parameter used to balance the bandwidth and reconstruction error of the modal components. Through iterative iteration, the optimal modal components are obtained. and center frequency .

[0007] Preferably, the spatial feature branch constructs a convolutional neural network model. The implementation process of the convolutional neural network model is as follows: the input data is sequentially passed through a stacked operation of "convolutional layer-ReLU activation-pooling layer" and finally outputs the result.

[0008] Preferably, a channel and spatial attention mechanism is introduced, including a channel attention module and a spatial attention module, for the input feature map. Perform the following operations: First, enter the channel attention module, and the formula for calculating channel weights is: in, For the input feature map, and These are average and max pooling, respectively. and These are the weight parameters of the fully connected layer. It is the ReLU activation function. For the Sigmoid function; Then, we enter the spatial attention module, where the formula for calculating spatial weights is: in, It is a convolutional layer. This indicates a channel splicing operation; The final output is: in, This indicates element-wise multiplication.

[0009] Preferably, the spatial attention module is replaced with a multi-scale dual-channel attention module, as follows: First, the channel refinement features are simultaneously fed into the identity mapping branch and multiple depthwise separable convolutional branches with different kernel sizes; Then, the output features of all branches are added and fused to integrate multi-scale information; Finally, the fused features are passed through a 1×1 convolutional layer, and the output is processed by a Sigmoid function to generate a channel attention weight map. Finally, the original input features are multiplied by the weight map channels to complete the feature reweighting.

[0010] Preferably, the bidirectional interactive fusion mechanism uses spatial feature tensors and temporal feature tensor As input, then perform the following operations: first, The query vector is obtained through linear projection. , The key vector is obtained through linear projection. Sum value vector .in, and Feature dimensions and Alignment of feature dimensions. Then... Query Generate attention weights and aggregate them. The temporal information in the output temporally enhanced spatial feature tensor is used to generate the output temporally enhanced spatial feature tensor. The calculation formula is as follows: in, , , It is a linear transformation matrix; then, The query vector is obtained through linear projection. , The key vector is obtained through linear projection. Sum value vector ; and Feature dimensions and Align the feature dimensions; then... Query Generate attention weights and aggregate them. The spatial information in the output spatially enhanced temporal feature tensor is used to generate the output spatially enhanced temporal feature tensor. The calculation formula is as follows: in, , , This is a linear transformation matrix; subsequently, the interaction results will be... , With the original spatiotemporal feature tensor , By concatenating features along the feature dimension, a fused composite feature is generated. As shown in the following formula: Then, an adaptive pooling layer is used to fuse the features. The features are compressed, flattened, and then input into a fully connected classifier to generate a prediction output.

[0011] Preferably, the spatial feature branch uses multi-level convolutional blocks for temporal compression and feature mapping, outputting a 128-dimensional feature vector.

[0012] Preferably, the temporal feature branch first reduces the length of the input sequence from 1024 steps to 128 steps through a feature recombination strategy, while expanding the feature dimension from 5 dimensions to 40 dimensions. The recombined data enters the Transformer encoder layer to learn temporal dependencies and apply Dropout regularization. Subsequently, the feature dimension is further expanded to 64 dimensions through two layers of BiLSTM.

[0013] A computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion as described above.

[0014] A processor for running a program, wherein the program executes the wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a dual-branch structure. One branch utilizes a CNN to extract spatial features, supplemented by an improved Multi-scale Depthwise Separable ConvolutionBlock Attention Module (MDSCBAM) to enhance the focusing ability on key features. The other branch mines long-term temporal dependencies using a Transformer model, introducing BiLSTM to compensate for its shortcomings in modeling local temporal dynamic characteristics, thus addressing the problem of insufficient utilization of temporal information in traditional fault diagnosis models. At the input data level, a combination of VMD and FFT preprocessing is used on the original vibration signal to improve the completeness of feature extraction. Finally, a bidirectional interactive fusion mechanism is used to integrate the spatiotemporal features extracted by the two branches, improving the performance and generalization ability of fault diagnosis. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0017] Figure 1 Here is a flowchart of the original signal preprocessing process; Figure 2 This is a schematic diagram of the CNN model structure; Figure 3 This is a schematic diagram of the CBAM module structure; Figure 4 This is a schematic diagram of the CAM model structure; Figure 5 This is a schematic diagram of the SAM model structure; Figure 6 This is a schematic diagram of the spatial attention module structure in MDSCBAM; Figure 7 This is a schematic diagram of the overall structure of MDSCBAM; Figure 8 A flowchart of the CNN-MDSCBAM workflow; Figure 9 This is a schematic diagram of the Transformer encoder structure; Figure 10 This is a schematic diagram of the BiLSTM network structure; Figure 11 This is a schematic diagram of the bidirectional interactive fusion mechanism (BIF). Figure 12 This is a schematic diagram of the CMTBB model structure; Figure 13 This is a schematic diagram showing the sensor installation location; Figure 14 This is a graph showing the performance of the NREL dataset model. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0020] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0022] In one embodiment of the present invention, a method for diagnosing wind turbine bearing faults based on dual-branch spatiotemporal feature fusion is provided, comprising the following steps: Step 1: Perform preprocessing operations on the input signal, which include at least variable mode decomposition and fast Fourier transform. Specifically: The original signal preprocessing process is as follows: Figure 1 As shown, the original vibration signal is decomposed into multiple modal components by VMD, and the original vibration signal is then used to generate frequency domain feature components by FFT. Finally, these components are vertically superimposed along the feature channel dimension to form the model input data.

[0023] Variational mode decomposition (VMD) is a fundamental non-recursive and adaptive signal decomposition technique that decomposes complex signals into several narrowband intrinsic mode functions (IMFs) with specific center frequencies. Compared to traditional mode decomposition, VMD effectively suppresses mode aliasing and endpoint effects, making it suitable for non-stationary signal analysis. VMD achieves signal decomposition by solving a constrained optimization problem, as follows: in, It is the first One modal component, It is the center frequency of the corresponding mode. It is the Dirac function. This represents the convolution operation. Represents the time derivative. This represents the original signal. By introducing the augmented Lagrangian function, the constrained problem is transformed into an unconstrained optimization problem. The augmented Lagrangian function is as follows: in, This is a regularization parameter used to balance the bandwidth and reconstruction error of the modal components. Through iterative iteration, the optimal modal components are obtained. and center frequency .

[0024] The Fast Fourier Transform (FFT) is an efficient algorithm for computing the Discrete Fourier Transform (DFT). The DFT is defined as follows: in, It is a time-domain signal. It is its corresponding frequency domain representation. This refers to the number of sampling points. The time complexity of directly calculating the DFT is... When the data size is large, the computational load increases dramatically, making it difficult to meet real-time requirements. FFT reduces computational complexity by using a divide-and-conquer strategy to break down the computation of DFT into several smaller subproblems.

[0025] While FFT can extract global spectral periodic features, it struggles to characterize the transient features of non-stationary vibration signals at different time scales. VMD, on the other hand, can adaptively decompose and acquire multi-scale time-domain components, fully preserving feature information at different time scales. Therefore, when processing the original signal, combining the multi-scale components of VMD with the global spectral features of FFT can generate comprehensive multi-scale feature information, fully capturing fault characteristics.

[0026] Step 2: Input the multi-scale signal tensor generated by the fusion of variational mode decomposition and fast Fourier transform, and output the probability distribution of fault categories using the fault diagnosis model. The fault diagnosis model includes spatial feature branches, temporal feature branches, and a fusion mechanism for the spatial feature branches and temporal feature branches. The spatial feature branches use multi-level convolutional blocks for temporal compression and feature mapping. The temporal feature branches include the encoder part of the Transformer and a bidirectional long short-term memory network. The fusion mechanism is specifically a bidirectional interactive fusion mechanism. The bidirectional interactive fusion mechanism generates query vectors to obtain information from the temporal features, or responds to the query of the temporal features as key values, forming a bidirectional information flow. Finally, the fusion interaction result is combined with the original spatiotemporal features.

[0027] In this embodiment, the structure of the convolutional neural network (CNN) model is as follows: Figure 2 As shown, the model employs a single-layer one-dimensional convolutional structure stacked with pooling layers for feature extraction, balancing the ability to capture local pattern information with computational efficiency. The entire process is as follows: input data sequentially passes through a stacked operation of "convolutional layer - ReLU activation - pooling layer," and finally outputs the result. This design reduces the computational cost and the number of parameters, achieves dimensionality reduction through pooling layers, and enhances robustness to small positional changes.

[0028] The structure of the Convolutional Block Attention Module (CBAM) is as follows: Figure 3 As shown, the CBAM structure includes a Channel Attention Module (CAM) and a Spatial Attention Module (SAM), the structures of which are as follows: Figure 4 and Figure 5 As shown: Input feature map First, enter CAM (Camera Management), and the formula for calculating channel weights is: in, For the input feature map, and These are average and max pooling, respectively. and These are the weight parameters of the fully connected layer. It is the ReLU activation function. This is the Sigmoid function. Then, we enter SAM, and the spatial weight formula is calculated as follows: in It is a convolutional layer; This indicates a channel concatenation operation; the final output is... ( (This indicates element-wise multiplication). This structure guides the network to focus on information-rich regions, improving performance.

[0029] Furthermore, this embodiment also provides an improved multi-scale depth-separable convolutional block attention module. The spatial attention module in the improved multi-scale depth-separable convolutional block attention module (MDSCBAM) structure is as follows: Figure 6 As shown.

[0030] Initially, the channel refinement features are simultaneously fed into the identity mapping branch and multiple depthwise separable convolutional branches using different kernel sizes (3, 5, and 7 to cover small, medium, and large receptive fields and extract multi-scale spatial features). Multi-scale depthwise separable convolutions extract multi-scale local spatial features while significantly reducing the number of parameters. Next, the output features of all branches are summed and fused to integrate multi-scale information. The fused features are then passed through a 1×1 convolutional layer to achieve information interaction and mixing between channels, enhancing feature expressiveness. The output is then processed by a sigmoid function to generate channel attention weight maps, which quantify the importance of each channel feature.

[0031] Finally, the original input features are multiplied by the weight map channels to complete feature reweighting, thereby enhancing the ability to extract key features.

[0032] Will Figure 6 The spatial attention module described in [the document] replaces the original spatial attention module in CBAM to construct MDSCBAM, as shown below. Figure 7 As shown, this module, based on CBAM, enables CNNs to simultaneously capture multi-scale spatial dependencies and further enhances their ability to identify key features, thereby improving the model's feature representation ability and generalization performance.

[0033] By embedding MDSCBAM immediately after each convolutional layer of a convolutional neural network (CNN), such as... Figure 8As shown, this mechanism achieves layer-by-layer dynamic feature optimization. Unlike the traditional approach of only adding attention modules at the end, this mechanism applies MDSCBAM early in feature extraction to enhance the representation of key features.

[0034] In the temporal feature branch, the encoder part of the Transformer is used, such as... Figure 9 As shown. The core structure of the Bidirectional Long Short-Term Memory (BiLSTM) network is as follows. Figure 10 As shown, it contains two independent LSTM layers: a forward layer and a backward layer. Each LSTM unit consists of three key gates: the forget gate controls the proportion of historical memory retained; the input gate adjusts the degree of integration of new information; and the output gate determines the output amount of the current state, calculated as follows: in, σ It is the sigmoid function. and It's about weights and biases. It is the previous hidden state. This is the current input. The final output is obtained by concatenating the hidden states of the forward and backward LSTMs: This provides rich contextual representations for downstream tasks.

[0035] In one embodiment of the present invention, to address the need for spatiotemporal feature extraction of vibration signals in bearing fault diagnosis, a dual-branch model is designed to simultaneously extract spatial and temporal features of the signal. Combined with a bidirectional interactive fusion (BIF) mechanism, feature fusion is achieved. The spatial features of the BIF can be used as query vectors. From the perspective of time series characteristics ( Information can be retrieved from ( ) or used as a key-value pair ( ) ) Query based on response time-series characteristics This forms a two-way information flow. Finally, the interaction results are integrated with the original spatiotemporal features to enhance the interaction representation capabilities while preserving the underlying information, thereby improving model performance.

[0036] like Figure 11 As shown, the bidirectional interactive fusion mechanism uses spatial feature tensors and temporal feature tensor As input. First, The query vector is obtained through linear projection. , The key vector is obtained through linear projection. Sum value vector .in, and Feature dimensions and Alignment of feature dimensions. Then... Query Generate attention weights and aggregate them. The temporal information in the output temporally enhanced spatial feature tensor is used to generate the output temporally enhanced spatial feature tensor. The calculation formula is as follows: in, , , It is a linear transformation matrix; then, The query vector is obtained through linear projection. , The key vector is obtained through linear projection. Sum value vector ; and Feature dimensions and Align the feature dimensions; then... Query Generate attention weights and aggregate them. The spatial information in the output spatially enhanced temporal feature tensor is used to generate the output spatially enhanced temporal feature tensor. The calculation formula is as follows: in, , , This is a linear transformation matrix; subsequently, the interaction results will be... , With the original spatiotemporal feature tensor , By concatenating features along the feature dimension, a fused composite feature is generated. As shown in formula (14): Then, an adaptive pooling layer is used to fuse the features. The features are compressed, flattened, and then input into a fully connected classifier to generate a prediction output.

[0037] This embodiment presents the overall structure of the fault diagnosis model, such as Figure 12 As shown in Table 1, the main structural parameters are as follows.

[0038] Table 1. CMTBB Model Structural Parameters The model input data is a multi-scale signal tensor generated by the fusion of variational mode decomposition and fast Fourier transform. The spatial feature branch uses multi-level convolutional blocks for temporal compression and feature mapping, outputting a 128-dimensional feature vector. The temporal feature branch first reduces the input sequence length from 1024 steps to 128 steps through a feature recombination strategy, while expanding the feature dimension from 5 dimensions to 40 dimensions. This operation aims to alleviate the computational redundancy and overfitting risk of the Transformer encoder layer when processing long sequences. The recombinated data enters the Transformer encoder layer, learns temporal dependencies, and applies Dropout regularization. Subsequently, two layers of BiLSTM are used to further expand the feature dimension to 64 dimensions. A bidirectional interactive fusion mechanism realizes spatiotemporal feature interaction, concatenating the output bi-branch enhanced features with the original features to form a 384-dimensional fused representation. After adaptive average pooling, the fault category probability distribution is output by a fully connected layer and a Softmax function.

[0039] This model combines multi-scale spatial features, temporal features, and a two-way interactive fusion mechanism with an adaptive alignment strategy for feature dimensions to mine spatiotemporal information in vibration signals and focus on key features, thereby improving the accuracy and robustness of fault diagnosis.

[0040] To further illustrate the implementation effect of this embodiment, a simulation is performed below based on the National Renewable Energy Laboratory (NREL) dataset. Data acquisition uses eight accelerometers placed at different locations, as shown below. Figure 13 The sampling frequency was 40 kHz. The bearing located at position AN6 suffered outer ring damage due to assembly errors; therefore, the vibration signal collected at this position was chosen as the research object. The data includes 450 samples each for normal and fault states, with each sample containing 1024 sampling points. All samples were randomly divided into training, validation, and test sets in a 6:2:2 ratio. Specific parameters of the samples are detailed in Table 2.

[0041] Table 2 Sample parameters of the NREL dataset like Figure 14 As shown in the accuracy vs. loss function curves on the validation set, the model converges rapidly around the 12th training epoch and remains highly stable after convergence, with the model loss value consistently at a low level. In the test set evaluation, its confusion matrix shows a 100% fault diagnosis accuracy, demonstrating excellent model generalization ability.

[0042] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0043] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.

[0044] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0045] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for fault diagnosis of wind turbine bearings based on dual-branch spatiotemporal feature fusion, characterized in that, Includes the following steps: S1: Perform preprocessing operations on the input signal, wherein the preprocessing operations include at least variational mode decomposition and fast Fourier transform; S2: Input the multi-scale signal tensor generated by the fusion of variational mode decomposition and fast Fourier transform, and output the probability distribution of fault categories using the fault diagnosis model; The fault diagnosis model includes a spatial feature branch, a temporal feature branch, and a fusion mechanism for the spatial feature branch and the temporal feature branch. The spatial feature branch uses multi-level convolutional blocks for temporal compression and feature mapping. The temporal feature branch includes the encoder part of the Transformer and a bidirectional long short-term memory network. The fusion mechanism is specifically a bidirectional interactive fusion mechanism. The bidirectional interactive fusion mechanism generates query vectors to obtain information from the temporal features, or responds to the query of the temporal features as key values, forming a bidirectional information flow. Finally, the fusion interaction result is combined with the original spatiotemporal features.

2. The wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion according to claim 1, wherein the variable mode decomposition is achieved by solving a constrained optimization problem to decompose the signal, characterized in that, By introducing the augmented Lagrangian function, the constrained problem is transformed into an unconstrained optimization problem. The augmented Lagrangian function is as follows: in, It is a regularization parameter used to balance the bandwidth and reconstruction error of the modal components. Through iterative iteration, the optimal modal components are obtained. and center frequency .

3. The wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion according to claim 1, characterized in that, The spatial feature branch constructs a convolutional neural network model. The implementation process of the convolutional neural network model is as follows: the input data passes through a stack of "convolutional layer-ReLU activation-pooling layer" in sequence, and finally outputs the result.

4. The wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion according to claim 3, characterized in that, Introducing channel and spatial attention mechanisms, including channel attention modules and spatial attention modules, for input feature maps Perform the following operations: First, enter the channel attention module, and the formula for calculating channel weights is: in, For the input feature map, and These are average and max pooling, respectively. and These are the weight parameters of the fully connected layer. It is the ReLU activation function. For the Sigmoid function; Then, we enter the spatial attention module, where the formula for calculating spatial weights is: in, It is a convolutional layer. This indicates a channel splicing operation; The final output is: in, This indicates element-wise multiplication.

5. The wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion according to claim 4, characterized in that, The spatial attention module is replaced with a multi-scale dual-channel attention module, as detailed below: First, the channel refinement features are simultaneously fed into the identity mapping branch and multiple depthwise separable convolutional branches with different kernel sizes; Then, the output features of all branches are added and fused to integrate multi-scale information; Finally, the fused features are passed through a 1×1 convolutional layer, and the output is processed by a Sigmoid function to generate a channel attention weight map. Finally, the original input features are multiplied by the weight map channels to complete the feature reweighting.

6. The wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion according to claim 1, characterized in that, The bidirectional interactive fusion mechanism uses spatial feature tensors and temporal feature tensor As input, then perform the following operations: first, The query vector is obtained through linear projection. , The key vector is obtained through linear projection. Sum value vector ,in, and Feature dimensions and Align the feature dimensions, and then... Query Generate attention weights and aggregate them. The temporal information in the output temporally enhanced spatial feature tensor is used to generate the output temporally enhanced spatial feature tensor. The calculation formula is as follows: in, , , It is a linear transformation matrix; then, The query vector is obtained through linear projection. , The key vector is obtained through linear projection. Sum value vector ; and Feature dimensions and Align the feature dimensions; then... Query Generate attention weights and aggregate them. The spatial information in the output spatially enhanced temporal feature tensor is used to generate the output spatially enhanced temporal feature tensor. The calculation formula is as follows: in, , , This is a linear transformation matrix; subsequently, the interaction results will be... , With the original spatiotemporal feature tensor , By concatenating features along the feature dimension, a fused composite feature is generated. As shown in the following formula: Then, an adaptive pooling layer is used to fuse the features. The features are compressed, flattened, and then input into a fully connected classifier to generate a prediction output.

7. The wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion according to claim 1, characterized in that, The spatial feature branch uses multi-level convolutional blocks for temporal compression and feature mapping, outputting a 128-dimensional feature vector.

8. The wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion according to claim 7, characterized in that, The temporal feature branch first reduces the length of the input sequence from 1024 steps to 128 steps through a feature recombination strategy, while expanding the feature dimension from 5 dimensions to 40 dimensions. The recombined data enters the Transformer encoder layer to learn temporal dependencies and apply Dropout regularization. Then, it continues to expand the feature dimension to 64 dimensions through two layers of BiLSTM.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion as described in any one of claims 1 to 8.

10. A processor, characterized in that, The processor is used to run a program, wherein the program executes the wind turbine bearing fault diagnosis method based on dual-branch spatiotemporal feature fusion as described in any one of claims 1 to 8.

Images