Fan bearing fault diagnosis method and system based on edge calculation

Abstract

The invention discloses a fan bearing fault diagnosis method and system based on edge calculation, and relates to the technical field of fault diagnosis of wind power generation. According to the method, an end-edge-training closed-loop architecture is adopted, and an original vibration signal is converted into a two-dimensional time-frequency spectrogram containing logarithmic power spectrum information. At an edge end, a lightweight student network integrating a double attention mechanism and a feature pyramid network is designed, and diagonal Mahalanobis distance is adopted to simplify measurement calculation. At a training end, a high-precision integrated teacher model containing a big nuclear attention and Transform cross attention mechanism is utilized, and complex knowledge is migrated to an edge student model through double distillation loss of output layer KL divergence and middle feature layer MSE. According to the method, the fan bearing fault can be quickly and accurately identified under the limited parameter quantity and calculation power.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for wind power generation, and specifically to a method and system for fault diagnosis of wind turbine bearings based on edge computing. Background Technology

[0002] Wind turbine generators operate in harsh environments for extended periods, and their bearings frequently fail, such as damage to the inner and outer rings and rolling elements of rolling bearings. Failure to diagnose these faults promptly can lead to generator downtime, increased maintenance costs, and even safety accidents. Traditional fault diagnosis methods primarily rely on signal processing techniques (such as Fourier transform and wavelet analysis) and shallow machine learning algorithms (such as support vector machines, SVM). While these methods can extract basic features, their ability to process complex nonlinear vibration signals is limited. Especially in noisy wind farm environments, their diagnostic accuracy is low, and they require extensive expert experience for feature engineering.

[0003] With the rise of deep learning technology, diagnostic methods based on convolutional neural networks (CNN), recurrent neural networks (RNN), or long short-term memory networks (LSTM) have gradually become mainstream. These methods can automatically learn high-dimensional features, improving diagnostic accuracy, but they have significant limitations: First, they heavily rely on massive amounts of labeled data and high-performance computing platforms for training and inference. For example, ResNet or Transformer-based models often have millions of parameters, making them difficult to deploy in actual wind turbine sites. Second, existing deep learning models face severe challenges in edge computing scenarios. Edge monitoring devices have limited computing resources and cannot directly run large models, requiring model compression to adapt. However, compression often leads to a significant decrease in accuracy and fails to maintain robustness under complex operating conditions. Furthermore, traditional methods often employ a centralized cloud architecture, uploading raw signals to remote servers for processing. This introduces data transmission latency, which cannot meet the real-time diagnostic needs of wind turbine bearing faults—bearing faults develop rapidly, and delays can lead to irreversible damage and economic losses.

[0004] Furthermore, while existing research on edge fault diagnosis attempts to introduce lightweight networks, several problems remain: First, generalization ability is weak, with model performance dropping sharply across different operating scenarios due to a mismatch between training data distribution and actual operational data. Second, processing speed is slow; while existing metric learning or attention mechanisms improve feature representation, computationally intensive operations at the edge lead to excessively high inference latency, preventing millisecond-level response times. Third, there is a lack of effective knowledge transfer mechanisms, hindering the efficient inheritance of complex knowledge from high-precision cloud models. Therefore, achieving lightweight, high-generalization fault diagnosis models and low-latency, high-precision identification at the edge is a pressing technical challenge. This invention addresses these limitations by proposing an innovative "edge-training" closed-loop architecture. Through a lightweight student network, double distillation, and an efficient metric mechanism, it significantly improves the efficiency and accuracy of edge diagnosis. Summary of the Invention

[0005] To address the challenges of obtaining fault samples for wind turbine bearings and the limited computing power of edge devices, which hinders rapid fault identification, this invention provides a wind turbine bearing fault diagnosis solution based on edge computing.

[0006] The technical solution of this invention is: a wind turbine bearing fault diagnosis method based on edge computing, comprising the following steps:

[0007] (1) Obtain the original vibration signal of the wind turbine bearing, generate a 64×64 two-dimensional time spectrum diagram through short-time Fourier transform and logarithmic power spectrum conversion, construct the signal characterization matrix, and standardize it using Min-Max normalization processing;

[0008] (2) Construct a dataset by sampling from the known sample set to generate a support set and a query set containing N categories and K samples for each category;

[0009] (3) Input the signal representation matrix into the edge student network and extract multi-scale intermediate features through the lightweight feature extraction backbone. The backbone adopts the AttentionMobileBlock module, which includes a depth-separable convolutional structure and integrates the SE module in channel attention and the spatial attention module based on large kernel convolution.

[0010] (4) The multi-scale intermediate features are fused using a feature pyramid network. The number of channels is unified by 1×1 convolution, and the feature fusion is performed from bottom to top by nearest neighbor interpolation to output an enhanced feature map.

[0011] (5) The similarity score between the query set features and the support set class prototype is calculated using the metric discrimination module. This module uses the diagonal Mahalanobis head and calculates the diagonal Mahalanobis distance by estimating the variance of each channel of the support set, avoiding the inversion operation of the complete covariance matrix, and outputs the fault diagnosis result.

[0012] (6) During the training phase, a teacher-student double distillation mechanism is adopted. The output layer distillation and feature-level distillation of the edge student network are performed using a preset high-precision integrated teacher model. The teacher model includes a multi-scale feature extraction module, a complete Mahalanobis distance branch, and a Transformer cross attention branch.

[0013] Furthermore, the logarithmic power spectrum conversion in step (1) includes: calculating the energy spectrum after STFT and performing logarithmic compression using the Power-to-dB function to enhance the identification of weak fault characteristics; the STFT is set with an FFT length of 512 and a frame shift of 512.

[0014] Furthermore, the lightweight feature extraction backbone of step (3) is composed of 4 stacked AttentionMobileBlock layers. Each module uses depthwise separable convolution, where the depthwise convolution kernel size is 3×3. Efficient feature extraction is achieved through PW expansion, DW convolution, dual attention enhancement and residual connection.

[0015] Furthermore, the output layer distillation in step (6) adopts KL divergence loss with a temperature coefficient T=4.0, combined with hard label loss with label smoothing; the feature-level distillation aligns the feature space resolution of the teacher and student models through bilinear interpolation, and calculates the mean squared error (MSE) loss after L2 normalization. The total loss function is defined as: L_total = α · L_distill + β · L_feature + (1-α-β) · L_hard.

[0016] Furthermore, the LKA branch of the high-precision integrated teacher model uses parallel 3×3, 5×5, and 7×7 convolutional kernels, combined with dilated convolutions to capture a super-large receptive field; the Transformer cross-attention branch uses Scaled Dot-Product Attention to calculate global relevance.

[0017] Furthermore, this method adopts a "end-edge-training" closed-loop architecture, deploying a lightweight student network and metric discrimination module at the edge, and a knowledge distillation module at the training end, supporting online updates and adaptations at the edge.

[0018] The above technical solution enables efficient edge computing in wind turbine bearing fault diagnosis. On one hand, by using time-spectrum graph transformation and a lightweight student network, robust features are extracted on resource-constrained edge devices. Combined with FPN fusion, multi-scale representation capabilities are enhanced, allowing for rapid identification of fault features in noisy windfield environments, avoiding the reliance on massive amounts of labeled data found in traditional methods. On the other hand, diagonal Mahalanobis distance simplifies metric calculations, significantly reducing inference latency and achieving millisecond-level response. A double-distillation mechanism transfers knowledge from a high-precision teacher model, ensuring that the student model's accuracy approaches that of large models with reduced parameters, making it suitable for complex fault scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the logic for wind turbine bearing fault diagnosis in this invention.

[0020] Figure 2 This is a flowchart of the method for diagnosing wind turbine bearing faults according to the present invention;

[0021] Figure 3 This is a model structure diagram of the AttentionMobileBlock module in the implementation of this invention. Detailed Implementation

[0022] Signal preprocessing and characterization construction

[0023] First, raw one-dimensional vibration signals were acquired using an accelerometer installed at the wind turbine bearing. The sampling frequency was set to 2048 Hz, and the signal acquisition duration was at least 10 seconds to ensure sufficient fault characteristics were captured. A sliding window was used to slice the signal, with a window length of 2048 sampling points to avoid information loss.

[0024] To convert a one-dimensional time-domain signal into a more discriminative two-dimensional representation, this embodiment employs the following steps for time-frequency conversion:

[0025] (1) Short-time Fourier Transform (STFT): Apply STFT to each window signal, set the FFT length to 512, the frame shift (hop_length) to 512, and the window function to Hann window.

[0026] (2) Logarithmic power spectrum conversion: The energy spectrum is obtained by squared the amplitude spectrum after STFT calculation, and then logarithmic compression is performed by applying the Power-to-dB function. This conversion enhances the identification of weak fault characteristics while suppressing the influence of background noise.

[0027] (3) Size standardization and normalization: The generated time-spectrum map is uniformly adjusted to a resolution of 64×64 pixels using bilinear interpolation. Then, Min-Max normalization is performed to map the pixel values ​​to the [0,1] interval, which is convenient for subsequent neural network input. This process constructs a signal representation matrix, which serves as the input data for the model.

[0028] In this embodiment, it is assumed that four types of bearing signals are collected: normal, inner ring fault, outer ring fault, and rolling element fault, with no less than 1,000 samples in each type, which will be used for subsequent dataset construction.

[0029] Edge Student Network Architecture

[0030] The lightweight StudentNet network, deployed at the edge, is designed for efficient inference on computationally limited devices. The total number of network parameters is kept to within approximately 50K, and the floating-point operation complexity does not exceed 100M. This network comprises the following core components.

[0031] (1) Feature extraction backbone: It consists of 4 stacked AttentionMobileBlock modules. The number of input channels in each module is 16, 32, 64, and 128 respectively. The structure of AttentionMobileBlock is as follows: Figure 3 As shown, the system includes: Point-Wise (PW) extended convolution: using 1×1 convolutions to expand the number of channels to four times the original number of channels. Depth-Wise (DW) convolution: using 3×3 kernels for depthwise separable convolution, with stride=2 for downsampling. Dual attention mechanism: Channel attention (SE module): calculating channel weights through global average pooling, with a compression ratio of 4 and the activation function being Sigmoid. Spatial attention module: based on 7×7 large kernel convolutions (decomposed into 1×7 and 7×1 convolutions to reduce computational cost), capturing local spatial correlations. Residual connection: adding the input to the attention-enhanced output to ensure stable gradient propagation. The output of each module is a multi-scale intermediate feature map with sizes of 32×32, 16×16, 8×8, and 4×4, respectively.

[0032] (2) Feature Pyramid Network (FPN): This network fuses intermediate features across multiple scales. First, 1×1 convolutions are used to unify the number of channels in each layer to 64. Then, nearest neighbor interpolation is used for bottom-up feature fusion: starting from the lowest resolution layer, features are upsampled and added to the features of the previous layer. Finally, an enhanced feature map with a size of 4×4×64 is output for subsequent metric discrimination.

[0033] Teacher Integrated Model Architecture

[0034] A high-precision ensemble teacher model was built on the training side for knowledge transfer. This model includes three parallel branches to improve generalization ability:

[0035] (1) Multi-scale feature extraction module: Parallel 3×3, 5×5, and 7×7 large kernel attention (LKA) modules are used. Each LKA module is combined with dilated convolution to capture a super-large receptive field. Branch outputs are fused by channel-level concat.

[0036] (2) Full Mahalanobis distance branch: Calculate the full covariance matrix based on the support set samples, and then calculate the Mahalanobis distance between the samples. Use Ledoit-Wolf regularization to avoid singular matrices.

[0037] (3) Transformer cross attention branch: Scaled Dot-Product Attention is used to calculate the global correlation between query samples and support samples.

[0038] Knowledge distillation training strategy

[0039] The training process adopts a closed-loop "end-edge-training" model. The dataset is expanded from the CWRU bearing fault dataset, including normal and three fault types, for a total of 4 classes (N=4). Each episode supports a set of K=5, and the query set has 15 samples.

[0040] (1) Output layer distillation: using KL divergence loss with a temperature coefficient T=4.0:

[0041] (2) Feature-level distillation: Align the feature space resolution of the teacher and student models by bilinear interpolation, and then calculate the mean square error after L2 normalization.

[0042] (3) Hard label loss: Cross-entropy loss combined with label smoothing (smoothing factor=0.1). Training uses the Adam optimizer with a learning rate of 1e-3 and a batch size of 4.

[0043] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this invention, and these modifications or substitutions should all be covered within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method and system for diagnosing wind turbine bearing faults based on edge computing, characterized in that... Includes the following steps: (1) Obtain the original vibration signal of the wind turbine bearing, generate a 64×64 two-dimensional time spectrum diagram through short-time Fourier transform (STFT) and logarithmic power spectrum conversion, and construct the signal characterization matrix; (2) Construct a dataset by sampling from the known sample set to generate a support set (Support) and a query set (Query) containing N categories and K samples for each category; (3) Input the signal representation matrix into the edge student network, extract multi-scale intermediate features through the lightweight feature extraction backbone, and fuse the multi-scale intermediate features using the feature pyramid network FPN to output an enhanced feature map; (4) Calculate the similarity score between the query set features and the support set class prototype using the metric discrimination module, and output the fault diagnosis result; (5) During the training phase, a teacher-student dual distillation mechanism is adopted, and the output layer distillation and feature-level distillation of the edge student network are performed using a preset high-precision integrated teacher model.

2. The wind turbine bearing fault diagnosis method based on edge computing according to claim 1, characterized in that, In step (1), the logarithmic power spectrum conversion includes: calculating the energy spectrum after STFT and performing logarithmic compression using the Power-to-dB function to enhance the identification of weak fault characteristics.

3. The wind turbine bearing fault diagnosis method based on edge computing according to claim 1, characterized in that, In step (3), the lightweight feature extraction backbone adopts AttentionMobileBlock, which includes a depthwise separable convolutional structure and integrates a channel attention (SE) module and a spatial attention module based on large kernel convolution.

4. The wind turbine bearing fault diagnosis method based on edge computing according to claim 1, characterized in that, In step (4), the metric discrimination module adopts the diagonal Mahalanobis Head, which calculates the diagonal Mahalanobis distance by estimating the variance of each channel of the support set, thus avoiding the inversion operation of the complete covariance matrix.

5. The wind turbine bearing fault diagnosis method based on edge computing according to claim 1, characterized in that, In step (5), the output layer distillation adopts KL divergence loss with temperature coefficient T and combines it with hard label loss with label smoothing; the feature level distillation aligns the feature space resolution of the teacher and student models through bilinear interpolation and calculates the mean square error (MSE) loss after L2 normalization.

6. The wind turbine bearing fault diagnosis method based on edge computing according to claim 1, characterized in that, The high-precision integrated teacher model includes: a multi-scale feature extraction module, which is composed of multiple large kernel attention LKA modules with different kernel sizes in parallel; a complete Mahalanobis distance branch, used to calculate the sample distance based on the complete covariance matrix; and a Transformer cross-attention branch, used to capture the global correlation between query samples and support samples.

7. A wind turbine bearing fault diagnosis system based on edge computing, characterized in that, include: The data preprocessing module is used to perform time-frequency conversion and logarithmic compression of the signal; The edge diagnostics module, deployed on edge computing nodes, includes a lightweight student network and a metric discriminator; the knowledge distillation module, deployed in the cloud or industrial control server, is used to guide the parameter optimization of the edge diagnostics module using an integrated teacher model.

Images