A wind power bearing fault diagnosis method based on decoupling-fusion generative adversarial network

Abstract

The application discloses a kind of wind power bearing fault diagnosis methods based on decoupling-fusion generative adversarial network, comprising: obtaining the vibration signal of wind power bearing under different fault conditions, and it is converted into two-dimensional time-frequency image;Build the generative adversarial network of double generator structure, wherein two generators are used to independently model time domain features and frequency domain features, realize the decoupling extraction of time-frequency characteristics;Respectively introduce axial attention mechanism in double generator, model along time axis and frequency axis respectively, to enhance the capture ability to local cross-dimensional correlation;Design cross-domain fusion module, fuse the time domain features and frequency domain features extracted by two generators, generate high-fidelity time-frequency image;The generated time-frequency image is used for training fault diagnosis model together with real time-frequency image, realize the classification and identification of wind power bearing faultThe application further improves the fault diagnosis performance.

Description

Technical Field

[0001] This invention relates to the field of wind power fault diagnosis technology, specifically to a wind power bearing fault diagnosis method based on decoupled-fused generative adversarial networks. Background Technology

[0002] In recent years, the rapid development of deep learning technology has fundamentally changed the research paradigm of mechanical condition monitoring and fault diagnosis. Data-driven methods can automatically learn discriminative features directly from raw sensor signals, eliminating the need for manually designed feature engineering. However, applying deep learning methods to wind turbine bearing fault diagnosis still faces a core bottleneck—the extreme scarcity of labeled fault samples. Due to safety risks and high costs, collecting real fault data from in-service wind turbines is significantly difficult. Generative Adversarial Networks (GANs) have been used to synthesize virtual fault samples, effectively expanding imbalanced datasets. However, several key scientific problems remain unsolved in GANs regarding fault sample generation.

[0003] First, existing GAN methods primarily employ a single generator structure, coupling time-domain and frequency-domain information within the same network. This limits the model's ability to simultaneously characterize fine-grained transient temporal features and structured frequency-domain patterns, while also leading to mutual interference between features. Second, traditional GANs mainly rely on convolutional operations with limited receptive fields. While effectively capturing local frequency-domain features, they are significantly insufficient in modeling the inherent long-range temporal dependencies of vibration signals. Therefore, although the samples generated by these methods can maintain spectral consistency, they struggle to accurately reflect the long-range temporal dependencies and periodic impact characteristics determined by the physical mechanisms of bearing failures. Consequently, diagnostic models trained on this synthetic data suffer from insufficient generalization ability and engineering reliability. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a wind turbine bearing fault diagnosis method based on decoupled-fused generative adversarial networks, thereby solving the problems existing in the background technology.

[0005] Technical solution: The wind turbine bearing fault diagnosis method based on decoupled-fused generative adversarial networks described in this invention includes the following steps:

[0006] Step 1: Acquire vibration signals of wind turbine bearings under different fault conditions and convert them into two-dimensional time-frequency images;

[0007] Step 2: Construct a generative adversarial network with a dual-generator structure, where the two generators are used to independently model the time-domain features and frequency-domain features, respectively, to achieve decoupled extraction of time-frequency features;

[0008] Step 3: Introduce axial attention mechanisms into the dual generators, modeling along the time axis and frequency axis respectively, to enhance the ability to capture local cross-dimensional correlations;

[0009] Step 4: Design a cross-domain fusion module to fuse the time-domain features and frequency-domain features extracted by the two generators to generate a high-fidelity time-frequency image;

[0010] Step 5: Use the generated time-frequency image and the real time-frequency image together to train the fault diagnosis model to achieve the classification and identification of wind turbine bearing faults.

[0011] Furthermore, in step 1, a two-dimensional time-frequency image is generated using the synchronous compressed wavelet transform method to improve the energy concentration of the time-frequency representation and the identifiability of fault features.

[0012] Furthermore, in step 2, the dual generator structure includes a time-domain generator and a frequency-domain generator, both of which use the same upsampling network structure. They independently model the time structure and the frequency structure through asymmetric convolution operations to achieve feature decoupling.

[0013] Furthermore, in step 2, the adversarial network is trained using the Wasserstein loss function with gradient penalty, and the generator and discriminator are optimized alternately. The generator is updated once after every few updates of the discriminator to ensure training stability.

[0014] Furthermore, in step 3, the axial attention mechanism decomposes the two-dimensional self-attention into two one-dimensional attention along the time axis and the frequency axis, respectively, to capture the temporal correlation within the same frequency band and the frequency dependence at the same moment, thereby enhancing the ability to model the time-frequency structure while reducing computational complexity.

[0015] Furthermore, in step 4, the cross-domain fusion module includes an intra-domain modeling submodule and a cross-domain interaction submodule. The intra-domain modeling submodule uses a state-space model to perform long-range dependency modeling of features along the time axis, enhancing the ability to express the temporal continuity of periodic impact patterns in vibration signals. The cross-domain interaction submodule realizes bidirectional information interaction between time-domain and frequency-domain features through an asymmetric convolution structure, and combines a gating mechanism to adaptively fuse the features of the two domains to generate a fused time-frequency representation.

[0016] Furthermore, the state-space model adopts a lightweight structure, models only along the time axis, and preserves the original feature information through residual connections, avoiding the additional computational overhead caused by modeling the frequency axis.

[0017] Furthermore, in the cross-domain interaction submodule, time-domain features are enhanced through frequency-to-time asymmetric convolution mapping, and frequency-domain features are enhanced through time-to-frequency asymmetric convolution mapping. The enhanced features are then used to generate soft attention weights through a gating mechanism to achieve adaptive modulation of the fused features.

[0018] The wind turbine bearing fault diagnosis system based on decoupled-fused generative adversarial networks described in this invention includes:

[0019] The signal acquisition and preprocessing module is used to acquire the vibration signal of the wind turbine bearing and convert it into a two-dimensional time-frequency image.

[0020] The dual generator module includes a time-domain generator and a frequency-domain generator, which are used to extract time-domain features and frequency-domain features, respectively.

[0021] The cross-domain fusion module is used to fuse the time-frequency features extracted by the two generators to generate a high-fidelity time-frequency image;

[0022] The fault diagnosis module is used to train a classification model based on the fused time-frequency images to identify fault types.

[0023] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: Addressing the limitations imposed by feature coupling in single-generator structures, this invention proposes a decoupling-fusion architecture with two complementary generators. These generators independently model features in the time and frequency domains, respectively, and combine asymmetric convolution and axial attention mechanisms to achieve refined feature extraction. To address the insufficient modeling of long-range temporal dependencies due to the limited receptive field of convolution, a gated cross-domain fusion module incorporating a state-space model is designed to achieve intra-domain temporal modeling, thereby efficiently capturing long-sequence correlations along the time axis and adaptively fusing complementary time-frequency representations to compensate for the shortcomings of traditional convolution operations in temporal modeling. Furthermore, the SDWGAN-GP method of this invention significantly outperforms existing mainstream methods in terms of sample generation quality and fault diagnosis performance under data scarcity conditions. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0026] like Figure 1 As shown, this embodiment of the invention provides a wind turbine bearing fault diagnosis method based on decoupled-fused generative adversarial networks, comprising the following steps:

[0027] Step 1: Collect vibration signals of wind turbine bearings under different fault conditions and convert them into a two-dimensional time-frequency representation using the SSQ-CWT (Synchronous Compressed Continuous Wavelet Transform) method. The specific process of SSQ-CWT is as follows: First, the original vibration signal... Morlet wavelet is selected as the mother wavelet. Calculate the continuous wavelet transform coefficients:

[0028]

[0029] in For scale parameters, For translation parameters, This is the complex conjugate of the mother wavelet. Then, the instantaneous frequency at each point is estimated based on the phase information of the wavelet coefficients:

[0030]

[0031] in The imaginary unit is used. Finally, a synchronous compression operation is performed to redistribute the wavelet coefficients on the original scale-time plane to their corresponding instantaneous frequency positions, resulting in the compressed time-frequency representation:

[0032]

[0033] in For the target frequency, This refers to the frequency resolution. For the first Each scale corresponds to a scale step increment.

[0034] Compared to traditional CWT (Continuous Wavelet Transform), SSQ-CWT significantly improves the energy concentration of the time-frequency representation through the aforementioned frequency redistribution, making the frequency components of bearing fault characteristics more clearly distinguishable. The obtained time-frequency representation is truncated to the target frequency range and adjusted to a fixed-size two-dimensional time-frequency image, which serves as the input for subsequent generative adversarial networks.

[0035] Step 2: Construct the SDWGAN-GP (Dedicated-Generator Wasserstein Generative Adversarial Network with Gradient Penalty Based on State-Space Model) model to generate corresponding time-frequency images for various faults. The dual-generator structure with axial attention decouples the time-domain and frequency-domain features, and the cross-domain fusion module based on SSMLite (Lightweight State-Space Model) integrates complementary information to synthesize high-fidelity time-frequency samples. The specific process is as follows:

[0036] SDWGAN-GP Model Architecture: The SDWGAN-GP proposed in this invention is a novel enhanced model based on the WGAN-GP (Wasserstein Generative Adversarial Network with Gradient Penalty) framework, which aims to more effectively characterize the time-frequency features of heterogeneity in vibration signals of engineering systems.

[0037] The model architecture includes: a dual generator with axial attention, a gated cross-domain fusion module based on SSMLite, and a discriminator.

[0038] During the decoupling phase, two generators targeting different domains model time-domain features and frequency-domain features respectively, and introduce axial attention mechanisms into their respective structures to capture local cross-dimensional correlations and enhance the expressive power of fine-grained spatial relationships.

[0039] In the fusion stage, SSMLite is introduced to model the long program sequence dependency along each axis, thereby enriching the time-frequency structure representation. The time features and spectral features are integrated through a gated cross-domain fusion module. High-fidelity time-frequency images are generated by using adaptive weighting and fusion convolution operations.

[0040] During the discrimination phase, real and synthetic samples are input into the discriminator to determine the real / synthetic labels, thereby enabling adversarial training against the generator and ultimately improving the quality of the generated samples.

[0041] Among them, the dual generator structure that introduces axial attention: In the traditional single generator architecture, time-frequency representation is usually modeled in a fully coupled manner, which to some extent limits the model's ability to simultaneously characterize fine-grained temporal transient features and structured frequency characteristics.

[0042] To improve the quality of generated images, this invention proposes a dual-generator architecture to extract time-frequency features in a decoupled manner. Two functionally specialized generators independently model the intrinsic temporal and spectral structures of the vibration signal using asymmetric convolution operations. Both generators employ a completely identical four-layer progressive upsampling structure: each upsampling layer consists of a deconvolution (transposed convolution) with a stride of 2, followed by convolutional feature refinement, batch normalization (BN), and ReLU activation. Starting from the latent vector, the generator progressively generates a 224×224 feature map through multiple levels of deconvolution operations.

[0043] To selectively capture domain-relevant dependencies, an Axial Attention module is introduced between the third and fourth layer upsampling modules of each generator. This mechanism decomposes traditional two-dimensional self-attention into two one-dimensional attentions along orthogonal axes, thus significantly reducing computational complexity. Specifically, given a feature map... Axial attention is divided into two phases, row attention (along the time axis) and column attention (along the frequency axis), which are executed sequentially.

[0044] For row attention, for each row of the feature map (Corresponding to a certain frequency band), through a linear projection matrix , , Calculate the query, key, and value vectors separately:

[0045]

[0046] Then calculate the row attention output:

[0047]

[0048] in The dimension of the key vector. Indicates the first The characteristic sequence of the row.

[0049] For column attention, for each column of the feature map (For a specific moment), attention is calculated along the frequency axis using the same calculation method:

[0050]

[0051] in Indicates the first The characteristic sequence of the column, , , This is the independent projection matrix for column attention. The results of row and column attention are sequentially superimposed onto the original feature map via residual connections:

[0052]

[0053] With a computational complexity of Unlike full two-dimensional attention, axial attention reduces complexity to This axis-based modeling design is highly compatible with the decoupling strategy of this invention: time axis attention is used to capture cross-time correlations within the same frequency band, while frequency axis attention characterizes cross-frequency dependencies at the same moment, thereby more accurately modeling time-frequency structural features.

[0054] A gated cross-domain fusion module based on SSMLite: To effectively fuse the complementary time-domain and frequency-domain representations learned by dual generators, this invention designs a gated cross-domain fusion module that achieves bidirectional information interaction through asymmetric convolution and gated attention mechanisms. Traditional convolution operations are limited by local receptive fields, making it difficult to model the long-range time-dependent characteristics commonly found in bearing vibration signals, such as periodic impact patterns spanning hundreds of time steps. To overcome this limitation, an SSMLite module is introduced to model intra-domain time dependencies before cross-domain feature fusion.

[0055] SSMLite is a variant of SSM (State-Space Model). The specific process of SSM is as follows: Given a feature map... ,in and Representing the frequency dimension and time dimension respectively, we reshape it into a time series form. That is, each sequence corresponds to a batch-band pair. Continuous-time state-space dynamics can be expressed as:

[0056]

[0057]

[0058] in Given the input sequence, Indicates a potential state. This is the output. To facilitate discrete-time computation, it is discretized using the zero-order hold (ZOH) method, resulting in:

[0059]

[0060]

[0061] in and These are the discretized state moments. For the current potential state, For the previous moment, Input the sequence for the current time. Output for the current time step. This is achieved by manipulating the matrix. By employing structured parameter design, SSMLite maintains linear computational complexity. At the same time, it enables efficient modeling of long-range dependencies, making it particularly suitable for time series modeling in time-frequency representation.

[0062] SSMLite is a lightweight, structured SSM that operates along the time axis and can efficiently capture long-range temporal dependencies.

[0063] Lite embodies two key design principles: SSM models only along the time axis, rather than simultaneously acting on both the time and frequency axes, thus significantly reducing computational overhead while focusing on the more critical time-dependent characteristics of vibration signals; it employs a shallow configuration (a two-layer structure). (This is sufficient to capture periodic patterns related to faults, provided that over-parameterization is avoided.)

[0064] Let the features from the time-domain generator and the frequency-domain generator be respectively and SSMLite models independently in two domains and introduces residual connections:

[0065]

[0066] in and The features are then modeled. Information exchange is then achieved through bidirectional cross-domain transformation. The time-to-frequency path employs an asymmetric convolution structure. The frequency-to-time path takes its symmetrical form ( ):

[0067]

[0068]

[0069] in and These represent cross-domain convolutional mappings that include batch normalization and ReLU activation, respectively. and Features are those resulting from information exchange.

[0070] To adaptively regulate cross-domain information flow, a spatially aware gating mechanism is introduced to generate a soft attention weight graph:

[0071]

[0072] in This represents the Sigmoid activation function. This indicates channel-level splicing. For gating network parameters, This means splitting the output vector into two parts. and These are the gating weights in the time and frequency domains, respectively. The final fused representation is obtained through gated modulation and output projection:

[0073]

[0074] in This represents element-wise multiplication. For the output projection module, there is one Convolutional thinning layer and one The channel compression layer consists of, This is the final fusion output. This fusion architecture can maintain temporal continuity while ensuring spectral structure consistency, thereby generating a time-frequency representation with high consistency and high fidelity.

[0075] Training strategy of SDWGAN-GP: Training is performed using the standard adversarial learning loss function of WGAN-GP.

[0076] The loss function of the discriminator is defined as:

[0077]

[0078] Where D represents the discriminator network; E represents the expected value of the output score; Represents the actual data distribution; Indicates the distribution of generated data; Indicates the generation of samples; Represents a random interpolated sample between the real sample and the generated sample; This represents the gradient of the discriminator output with respect to the interpolated samples; is the gradient penalty term. This is used to constrain the discriminator to satisfy the 1-Lipschitz continuity condition; This represents the gradient penalty weight coefficient.

[0079] The generator loss function is defined as:

[0080]

[0081] Here, G represents the generator network, whose optimization objective is to maximize the discriminator D's performance on the generated samples. The output score makes the generated data distribution... Approximating the true data distribution .

[0082] The dual generator and discriminator (critic) are optimized alternately, with the generator's parameters updated every three updates to the discriminator. The training process includes the following steps:

[0083] First, set the hyperparameters required for training, including the maximum number of training epochs. Batch size Learning rate Number of discriminator iterations in each round Gradient penalty coefficient The optimizer uses Adam, with its momentum parameter set to... , Then initialize the time domain generator parameters. Frequency domain generator parameters Cross-domain fusion parameters and discriminator parameters .

[0084] In each training round, first execute The discriminator is updated every time. Each update involves analyzing the distribution of real time-frequency image data. A batch of samples were collected. At the same time, from the standard normal distribution A batch of potential noise vectors were sampled. The noise vector is input into the time-domain generator and the frequency-domain generator, respectively. Each of them has its own feature representation. and Then through the cross-domain integration module Generate fused pseudo-samples Then from a uniform distribution Medium-sample interpolation coefficients to construct interpolation samples And calculate the discriminator loss including the gradient penalty term:

[0085] ;

[0086] in, This represents the discriminator network; This represents the expected value of the output score. Represents the actual data distribution; Indicates the distribution of generated data; Indicates the generation of samples; Represents a random interpolated sample between the real sample and the generated sample; This represents the gradient of the discriminator output with respect to the interpolated sample; This is a gradient penalty term used to constrain the discriminator to satisfy the 1-Lipschitz continuity condition; This represents the gradient penalty weight coefficient. Based on this loss, the discriminator parameters are updated via gradient descent: .

[0087] Finish After the discriminator is updated, a generator update is performed. A new batch of noise vectors is then sampled. Pseudo-samples are generated using a dual generator and a fusion module, and the generator loss is calculated:

[0088]

[0089] Among them, among them, This represents a generator network whose optimization objective is to maximize the discriminator. For generated samples The output score makes the generated data distribution... Approximating the true data distribution Update the generator parameters accordingly: Repeat the above process until the maximum number of training rounds is reached. The final output is the trained generator. , and discriminator .

[0090] (3) The generated time-frequency images and the real time-frequency images are used together to train the fault diagnosis model to realize the classification and identification of wind turbine bearing faults.

Claims

1. A method for fault diagnosis of wind turbine bearings based on decoupled-fused generative adversarial networks, characterized in that, Includes the following steps: Step 1: Acquire vibration signals of wind turbine bearings under different fault conditions and convert them into two-dimensional time-frequency images; Step 2: Construct a generative adversarial network with a dual-generator structure. The two generators independently model the time-domain and frequency-domain features, respectively, achieving decoupled extraction of time-frequency features. The dual-generator structure includes one time-domain generator and one frequency-domain generator, both using the same upsampling network structure. They independently model the time structure and spectral structure through asymmetric convolution operations, achieving feature decoupling. The adversarial network is trained using a Wasserstein loss function with gradient penalty. The generator and discriminator are optimized alternately, with the generator updated after every few updates to the discriminator, ensuring training stability. Step 3: Introduce axial attention mechanisms in the dual generators to model along the time axis and frequency axis respectively, so as to enhance the ability to capture local cross-dimensional correlations. The axial attention mechanism decomposes the two-dimensional self-attention into two one-dimensional attentions along the time axis and frequency axis, which are used to capture the temporal correlation within the same frequency band and the frequency dependence at the same time, respectively, thereby enhancing the ability to model the time-frequency structure while reducing computational complexity. Step 4: Design a cross-domain fusion module to fuse the time-domain features and frequency-domain features extracted by the two generators to generate a high-fidelity time-frequency image; Step 5: Use the generated time-frequency image and the real time-frequency image together to train the fault diagnosis model to achieve the classification and identification of wind turbine bearing faults.

2. The wind turbine bearing fault diagnosis method based on decoupled-fused generative adversarial networks according to claim 1, characterized in that, In step 1, a two-dimensional time-frequency image is generated using the synchronous compressed wavelet transform method.

3. The wind turbine bearing fault diagnosis method based on decoupled-fused generative adversarial networks according to claim 1, characterized in that, In step 4, the cross-domain fusion module includes an intra-domain modeling submodule and a cross-domain interaction submodule. The intra-domain modeling submodule uses a state-space model to perform long-range dependency modeling of features along the time axis, enhancing the ability to express the temporal continuity of periodic impact patterns in vibration signals. The cross-domain interaction submodule realizes bidirectional information interaction between time-domain and frequency-domain features through an asymmetric convolution structure, and combines a gating mechanism to adaptively fuse features from both domains to generate a fused time-frequency representation.

4. The wind turbine bearing fault diagnosis method based on decoupled-fused generative adversarial networks according to claim 3, characterized in that, The state-space model adopts a lightweight structure, models only along the time axis, and preserves the original feature information through residual connections, avoiding the additional computational overhead of modeling the frequency axis.

5. The wind turbine bearing fault diagnosis method based on decoupled-fused generative adversarial networks according to claim 3, characterized in that, In the cross-domain interaction submodule, time-domain features are enhanced through frequency-to-time asymmetric convolution mapping, and frequency-domain features are enhanced through time-to-frequency asymmetric convolution mapping. The enhanced features are then used to generate soft attention weights through a gating mechanism to achieve adaptive modulation of the fused features.

6. A wind turbine bearing fault diagnosis system based on decoupled-fused generative adversarial networks, characterized in that, The method described in any one of claims 1-5 is used to implement the method, comprising: The signal acquisition and preprocessing module is used to acquire the vibration signal of the wind turbine bearing and convert it into a two-dimensional time-frequency image. The dual generator module includes a time-domain generator and a frequency-domain generator, which are used to extract time-domain features and frequency-domain features, respectively. The cross-domain fusion module is used to fuse the time-frequency features extracted by the two generators to generate a high-fidelity time-frequency image; The fault diagnosis module is used to train a classification model based on the fused time-frequency images to identify fault types.

Images