Aerospace bearing fault diagnosis method based on GADF-driven KAN-Swin Transformer double-branch network

Abstract

The invention relates to the technical field of aerospace bearing fault diagnosis methods, and particularly discloses an aerospace bearing fault diagnosis method based on a GADF-driven KAN-Swin Transform double-branch network, and the method comprises the following steps: 1, signal collection and GADF image generation: collecting bearing vibration signals in different health states through a channel sensor, and generating a GADF image; the method comprises the following steps of: performing window division on original time sequence data to construct a training / testing data set, converting each signal window into a two-dimensional image by adopting a Grami angular difference field method, reserving time sequence correlation characteristics of the two-dimensional image through polar coordinate transformation, normalizing the generated GADF image, storing the normalized GADF image as an RGB (Red, Green, Blue) format data set, and dividing a training set, a verification set and a testing set according to a ratio of 7: 2: 1. According to the aerospace bearing fault diagnosis method based on the GADF-driven KAN-Swin Transform double-branch network, a reliable solution is provided for intelligent health monitoring of the aerospace bearing through collaborative integration of mechanism fusion and an attention architecture.

Description

Technical Field

[0001] The present invention relates to the technical field of aerospace bearing fault diagnosis methods, and specifically to an aerospace bearing fault diagnosis method based on a dual-branch network of KAN-Swin Transformer driven by GADF. Background Technique

[0002] Traditional fault diagnosis methods mainly rely on time-domain, frequency-domain, and time-frequency domain signal processing techniques. However, in practical engineering applications, the diagnostic effects of these methods are often restricted by noise interference and working condition changes. To break through these limitations, deep learning-based methods have gradually become a research hotspot. Deep learning models such as convolutional neural networks (CNNs), generative adversarial networks (GANs), and Transformers have demonstrated powerful capabilities in image feature extraction and fault pattern recognition. For example, Zhang et al. proposed a fault diagnosis method based on a deep convolutional neural network, significantly improving the recognition accuracy and anti-noise performance of rotating machinery faults. Guo et al. achieved cross-domain fault diagnosis by combining transfer learning with a deep CNN. Wang and Li used GANs for fault sample enhancement and feature extraction to solve the problem of small sample learning. In addition, Chen combined an improved time-frequency analysis method with a CNN to achieve refined diagnosis of rolling bearing faults. Transformers and attention mechanisms have also become key tools for improving fault diagnosis accuracy. The Swin Transformer proposed by Fu et al. demonstrated excellent performance in image classification and fault diagnosis through hierarchical window design and window shifting strategies. Wu et al. reviewed the application of deep learning in mechanical fault diagnosis and discussed the advantages of attention mechanisms in extracting key fault features;

[0003] Although existing methods have improved the accuracy and robustness of fault diagnosis to a certain extent, there are still several challenges in aerospace bearing diagnosis: First, most existing models lack effective feature representation capabilities under noise interference and variable working conditions, resulting in fuzzy or indistinguishable fault features, thus affecting the diagnostic accuracy. Although methods such as CNNs and Transformers have been applied in this field, their performance in complex environments still has limitations. Second, existing local feature extraction models are insufficient in capturing the global structural dependence relationships within signals, hindering a comprehensive understanding of hidden information. Summary of the Invention

[0004] To achieve the above objectives, the present invention is realized through the following technical solutions: An aerospace bearing fault diagnosis method based on a dual-branch network of KAN-Swin Transformer driven by GADF, comprising the following steps:

[0005] Step 1: Signal acquisition and GADF image generation. The bearing vibration signals under different health states are collected through channel sensors. The original time-series data is windowed to construct a training / test dataset. The Gramian Angular Difference Field (GADF) method is used to convert each signal window into a two-dimensional image, and then polar coordinate transformation is performed to retain its time-series correlation features. The generated GADF images are normalized and stored as an RGB format dataset, and the training set, validation set, and test set are divided according to a 7:2:1 ratio.

[0006] Step 2: Hybrid feature extraction and model training. The training set of GADF images is input into the KANSwinT-GAM architecture for joint feature learning, and the model processes through a dual-branch parallel structure.

[0007] Step 3: Fault diagnosis and verification. The test set of GADF images is input into the trained model for health state recognition. The KAN-Swin Transformer and GAM frameworks use a multi-level feature fusion mechanism, combining the structure-aware ability of the KAN-Swin Transformer and the texture-sensitive characteristics of the CNN-GAM branch to achieve collaborative extraction of fault features.

[0008] Preferably, the size of the GADF image in step 1 is 224×224.

[0009] Preferably, the dual-branch parallel processing in step 2 is as follows:

[0010] KAN-Swin Transformer branch: Capturing local spatial dependencies based on the shifted window self-attention mechanism, processing 4×4 non-overlapping image patches using a hierarchical structure, using the KAN network instead of the traditional MLP for classification, and optimizing the non-linear mapping of features through a dynamic B-spline grid to generate multi-scale features.

[0011] CNN-GAM branch: Extracting global texture features through a CNN backbone network that stacks Conv2D, ReLU, and GAM attention layers, where the GAM mechanism enhances discriminative features through the concatenation of channel attention and spatial attention.

[0012] The hierarchical features output by the KAN-Swin Transformer branch and the high-dimensional global features of the CNN-GAM branch are cross-modally fused through channel concatenation and adaptive average pooling.

[0013] It has the following beneficial effects:

[0014] The aerospace bearing fault diagnosis method based on the GADF-driven KAN-Swin Transformer dual-branch network first encodes one-dimensional vibration signals into two-dimensional time-frequency images through the Gramian angular field, and uses polar coordinate mapping to retain the temporal dependence and spectral dynamic characteristics, overcoming the limitation of traditional time-frequency analysis methods being sensitive to noise. Furthermore, it innovatively uses B-spline basis functions to replace the traditional multi-layer perceptron to construct the KAN-Swin Transformer module, and enhances the non-linear mapping ability through a dynamic grid adjustment strategy, effectively modeling transient shocks and periodic patterns while reducing the parameter complexity. Further, a dual-branch parallel architecture is proposed: the KAN-Swin Transformer branch extracts local structure features through a hierarchical window attention mechanism, while the CNN-GAM branch enhances global texture perception through multi-scale convolution and channel-spatial attention fusion. Finally, through cross-modal feature splicing and adaptive pooling, the collaborative optimization of global-local characteristics is achieved, significantly improving the discriminability of fault patterns in complex noise environments. The proposed method achieves 100% classification accuracy on two different aerospace bearing datasets. Through ablation experiments, the synergistic effect of the KAN module, Swin Transformer, and CNN-GAM is verified. Compared with the baseline model, it shows a better cross-rotational speed condition recognition rate, and the robustness of the framework to noise and variable load conditions is confirmed. Through the collaborative integration of mechanism fusion and attention architecture, this framework provides a reliable solution for the intelligent health monitoring of aerospace bearings. Description of the Drawings

[0015] Figure 1 This is the structure diagram of KAN of the present invention;

[0016] Figure 2 This is the structure diagram of the KAN-Swin Transformer module of the present invention;

[0017] Figure 3 This is the dual-branch network structure of KAN-Swin Transformer of the present invention;

[0018] Figure 4 This is the confusion matrix and dataset diagram of the method proposed by the present invention under different datasets;

[0019] Figure 5 This is the visualization diagram of the T-distributed stochastic neighbor embedding of the extracted features corresponding to the model proposed by the present invention under two datasets;

[0020] Figure 6 This is the comparison diagram of the accuracy, precision, and F1 scores of six different diagnostic models in datasets A and B of the present invention. Detailed Implementation Modes

[0021] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] Please refer to Figures 1-6 , the present invention provides a technical solution: an aerospace bearing fault diagnosis method based on a GADF-driven KAN-SwinTransformer dual-branch network, including the following steps:

[0023] Step 1: Signal acquisition and GADF image generation. Collect the bearing vibration signals in different health states through a channel sensor, divide the original time-series data into windows to construct a training / test dataset, use the Gramian angular difference field method to convert each signal window into a two-dimensional image, and then retain its time-series correlation features through polar coordinate transformation to generate a GADF image with a size of 224×224. After normalization, it is stored as an RGB format dataset, and the training set, validation set, and test set are divided according to the ratio of 7:2:1;

[0024] Step 2: Hybrid feature extraction and model training. Input the training set GADF images into the KANSwinT-GAM architecture for joint feature learning. The model processes through two branches in parallel:

[0025] KAN-Swin Transformer branch: Capture local spatial dependencies based on the shifted window self-attention mechanism, process 4×4 non-overlapping image patches using a hierarchical structure, use the KAN network instead of the traditional MLP for classification, and optimize the non-linear mapping of features through a dynamic B-spline grid to generate multi-scale features;

[0026] CNN-GAM branch: Extract global texture features through a CNN backbone network that stacks Conv2D, ReLU, and GAM attention layers. Among them, the GAM mechanism enhances discriminative features through the concatenation of channel attention and spatial attention;

[0027] The hierarchical features output by the KAN-Swin Transformer branch and the high-dimensional global features of the CNN-GAM branch are cross-modally fused through channel concatenation and adaptive average pooling;

[0028] Step 3: Fault diagnosis and verification. Input the test set GADF images into the trained model for health state recognition. The KAN-Swin Transformer and GAM framework achieve the collaborative extraction of fault features through a multi-level feature fusion mechanism, combining the structure perception ability of KAN-Swin Transformer and the texture sensitivity characteristics of the CNN-GAM branch.

[0029] The traditional Swin Transformer model realizes the global modeling of image features through the window multi-head self-attention and shifted window mechanisms. However, there are still two key problems in its multi-layer perceptron module, abbreviated as the MLP module: First, the parameter quantity of the fully connected structure of multiple MLPs grows with the square of the feature dimension, which is prone to overfitting when the training data is limited. Second, the linear combination of fixed activation functions limits the nonlinear representation ability of the model. To overcome the above defects, the present invention proposes an improved KAN-Swin Transformer module based on the Kolmogorov-Arnold theorem, and constructs a more expressive nonlinear mapping structure by replacing the traditional MLP with a learnable B-spline basis function.

[0030] According to the Kolmogorov-Arnold theorem, any multivariate continuous function can be decomposed into a nested superposition of a finite number of univariate functions, and this theorem can be expressed by the following formula:

[0031]

[0032] where is the outer function, is the inner univariate function. Inspired by this, the basis function is parameterized as a B-spline curve, which is expressed by the following formula:

[0033]

[0034] In the formula is the basis spline function, is the trainable coefficient. By explicitly modeling the function decomposition process as a neural network layer, a Kolmogorov-Arnold network, abbreviated as KAN, is formed. In KAN, each layer is defined as , The values of are determined by the corresponding input and output of this layer. The structure of KAN is as Figure 1 shown, Figure 1 There are a total of n features as inputs. The input layer first introduces these features into the hidden layer. For each input xp, through m inner functions mapping: , where m is the number of hidden layer nodes. The hidden layer generates an intermediate representation by applying a spline function to the input features and then passes it to the output layer. The output hq of the hidden layer is combined through an outer function to obtain the final output: .

[0035] , Spline functions are usually applied.

[0036] In the standard Swin Transformer Block, the MLP module is responsible for performing non-linear transformations on the features output by self-attention. The KAN-Swin Transformer module proposed in the present invention replaces it with a KAN structure. On the basis of the standard Swin Transformer Block, the KAN-Swin Transformer module forms a double-residual cascade structure by introducing the Kolmogorov-Arnold network, abbreviated as KAN, and reconstructing the feed-forward layer, as Figure 2 shown. Its mathematical expression can be divided into two feature transformation stages:

[0037] Window self-attention transformation, which retains the original window self-attention mechanism. The output features are input to the KAN layer after being normalized by LayerNorm. This process is represented by the following formula:

[0038]

[0039] where represents window multi-head self-attention, is the layer normalization operation, is the input feature, where B is the batch size, N = H × W is the number of spatial positions, and d is the feature dimension.

[0040] KAN feed-forward transformation. This process is represented by the following formula, and a two-stage KAN structure is used to replace the traditional MLP:

[0041]

[0042] In the formula is the improved Kolmogorov-Arnold network layer. For the input , the output process of a single-layer KAN is represented by the following formula:

[0043]

[0044] The output of the entire KAN network is represented by the following formula:

[0045]

[0046] Among them, is the basic activation function, is the basis function coefficient matrix, is the B-spline basis function of the i-th input feature, is the spline weight, and the output of window self-attention is added to the input through a skip connection to retain the original feature information, and the output of the KAN layer is added to to form a cascaded residual structure. The window self-attention captures spatial local correlations, and the KAN layer models global frequency domain features through basis function combinations to form a complement. The double residual structure improves the gradient straight-through rate, which is higher than that of the original Swin Transformer, alleviating the deep network degradation problem.

[0047] In the KAN network, by introducing a dynamic grid adjustment strategy, the spline nodes are optimized according to the input feature distribution, and the calculation formula is as follows:

[0048]

[0049] Among them is the uniform grid, is the adaptive grid calculated based on feature quantiles, is the mixing coefficient.

[0050] The residual connection structure ensures the effective backpropagation of gradients. Combined with the dynamic grid update mechanism, it alleviates the problem of gradient disappearance in deep network training. Compared with the parameter quantity of the traditional MLP, KAN reduces the parameter complexity to .

[0051] The overall architecture of the proposed model is as shown in Figure 3 . It consists of a GADF preprocessing module and a KAN-Swin Transformer and a CNN-GAM dual-branch network. The GADF converts the one-dimensional vibration signal into a two-dimensional time-frequency image through the Gramian angular field, capturing the periodicity and transient impact characteristics of the signal. The KAN-Swin Transformer and GAM adopt a dual-branch parallel structure. Among them, the KAN-Swin Transformer branch extracts local texture features through hierarchical window attention, and the CNN-GAM branch enhances the spatial representation ability using multi-scale convolution and global attention. The features of the two branches are fused through cross-modal fusion to achieve high-precision fault classification. The complete structure of the model is as shown in Figure 3 . The parameter configurations of each layer are shown in the following table:

[0052]

[0053] a) Time-frequency feature encoding;

[0054] The GADF module converts the one-dimensional vibration signal into a two-dimensional time-frequency image through the Gramian Angular Difference Field. Given the sampled signal , GADF constructs a Gramian matrix through polar coordinate mapping to generate an RGB image with a size of 224×224. This process converts the time dependence of the signal into spatial texture features, where the diagonal of the image represents the autocorrelation of the time series, and the off-diagonal elements reflect the phase relationship across time steps, effectively retaining the joint time-frequency features of the signal.

[0055] b) Dual-branch feature extraction;

[0056] The model adopts a dual-branch parallel structure to achieve multi-scale feature extraction:

[0057] KAN-Swin Transformer backbone network: Based on the window attention mechanism of the Swin Transformer improved by KAN, it calculates spatial correlations within a 7×7 local window using 4 heads of attention, adopts third-order B-spline basis functions, and dynamically adjusts the node positions according to the feature distribution to implement the dynamic KAN layer. The core innovation lies in replacing the traditional MLP layer with the KAN module, and enhancing the non-linear fitting ability through adaptive grid adjustment.

[0058] GAM-CNN branch: Construct a deep convolutional network, and introduce a GAM attention module in each layer. This module adopts a dual attention mechanism: Channel attention learns channel weights through a fully connected layer , and spatial attention generates a spatial mask through a 7×7 convolution (Spatial pyramid pooling with a 7×7 convolution kernel), and finally realizes the output to achieve double feature enhancement.

[0059] c) Multi-modal feature fusion;

[0060] The local features of the KAN-Swin Transformer and the global features of the GAM-CNN are fused into joint features through channel concatenation. Adaptive average pooling is used to compress the spatial dimension and retain the most discriminative channel responses. Finally, the fault state classification is achieved through a fully connected layer. The loss function uses cross-entropy loss. This design combines the long-range dependence modeling ability of the KAN-Swin Transformer and the local feature extraction advantages of the CNN, and strengthens the key features through the attention mechanism, significantly improving the classification robustness under complex working conditions.

[0061] The present invention uses data under the sampling frequency of 51.2KHZ, the shaft rotation speed of 100HZ (6000rpm), and no-load conditions. The dataset includes three health states: inner race fault, roller fault, and normal state. According to the different degrees of faults, the inner race and roller faults are each divided into 3 degrees. Therefore, this dataset has a total of 7 states. Each fault type has acceleration information of 6 channels, and each channel has 512000 data points. 250 samples are selected for each working condition to form a sample set. The division of the dataset is consistent with the method of Dataset A. The specific content is shown in the following table:

[0062]

[0063] The proposed method has achieved good results in both datasets. Next, it will be analyzed from two aspects: the diagnostic result analysis of the proposed method and the T-SNE visualization result analysis:

[0064] a) Diagnostic results based on the framework of the proposed method

[0065] To prove the effectiveness of the proposed method, the average accuracy of the proposed model is obtained after five runs on two datasets. Under two different datasets, the fault recognition rate of the proposed model always exceeds 99.9%, indicating that the method can effectively distinguish bearing samples with different health conditions under complex working conditions. To further verify the discrimination ability of the proposed model, a confusion matrix analysis is performed on the two datasets. Figure 4 (a) and (b) respectively show the normalized confusion matrices of Dataset A and Dataset B. The diagonal elements represent the classification accuracy of each category, and the non-diagonal elements reflect the misjudgment situation.

[0066] For Dataset A, the model achieves 100% accuracy in all four health states. It is worth noting that there is no misjudgment between similar fault degrees, indicating the sensitivity of the model to the transient impact characteristics in the GADF image. For Dataset B, the model maintains perfect classification in seven working conditions, including the fine distinction of different defect sizes between inner race faults and roller faults, verifying the robustness of the method in fault classification under complex working conditions.

[0067] b) T-SNE visualization result analysis

[0068] To further observe the dispersion and clustering patterns of different samples, Figure 5 shows the visualization of the T-distributed Stochastic Neighbor Embedding of the extracted features corresponding to the proposed model under two datasets. It can be found from Figure 5 that as the network deepens, similar fault features gradually cluster. Figure 5(a,b) shows the local features of two datasets extracted by KAN-Swin Transformer. The feature distributions of samples from different classes present a multi-cluster structure, with continuous transition regions. Figure 5 (c,d) shows the global features of two datasets extracted by GAM. The positions of samples from different classes are close and there is a small amount of overlap, as Figure 5 shown in (e,f). The intra-class clustering and inter-class dispersion of all samples have been improved, which is consistent with our expectation. This is mainly because the dual-branch network with KAN-Swin Transformer structure enhances the differences between different faults, and similar faults are clustered together, thus obtaining satisfactory classification results.

[0069] To further examine the contributions of KAN, Swin Transformer, and CNN-GAM modules in the proposed model, ablation experiments were conducted. The specific adjustment plans for ablation are as follows:

[0070] 1) CNN-Transformer: The basic Transformer model without other modules, used as a comparison benchmark;

[0071] 2) CNN-Swin Transformer: Only contains the Swin Transformer module, testing the effect of multi-scale feature extraction of the Swin Transformer module;

[0072] 3) KAN-Swin Transformer: Contains KAN and Swin Transformer modules, removes the CNN-GAM module, used to integrate KAN into the Swin Transformer module combination to improve the model performance;

[0073] 4) CNN-GAM: Removes KAN and Swin Transformer modules, only retains the CNN-GAM module, used to observe the contribution of the CNN-GAM module to feature extraction and model performance;

[0074] 5) Swin Transformer-GAM: Contains Swin Transformer and CNN-GAM modules, removes the KAN module, used to evaluate the impact of the Swin Transformer and CNN-GAM combination on the model performance;

[0075] 6) Dual-branch network with KAN-Swin Transformer: Convert the one-dimensional signal into a two-dimensional time-frequency image GADF, and test the synergistic effect of each module through a combination of KAN, Swin Transformer, and CNN-GAM modules;

[0076] The ablation experiment results on the two datasets are shown in the following table. The table lists different metrics for the two datasets. First, overall, the proposed method consistently outperforms all other methods in various metrics, indicating its superiority. Second, compared with Swin Transformer-GAM, the recognition rate of the proposed method has increased, thanks to the reasoning ability introduced by the KAN structure in the classification decision-making process.

[0077]

[0078] The ablation experiment results on the two datasets are as Figure 6 shown. Through the ablation experiment, it is found that the joint design of the KAN-SwinTransformer module and the CNN-GAM dual-branch architecture significantly improves the model performance. In the KAN-SwinTransformer module, the B-spline-based Kolmogorov-Arnold network replaces the MLP layer of the standard Swin Transformer, and through dynamic adjustment of the activation function and the adaptive grid mechanism, it enhances the non-linear modeling ability for high-frequency transient impact features. Compared with the traditional Swin Transformer, KAN-Swin Transformer achieves an accuracy of 98.10% on dataset A while reducing the number of parameters, verifying its efficiency and parameter efficiency.

[0079] The double-branch architecture has achieved a significant performance leap through global-local feature complementarity. The KAN-Swin Transformer branch utilizes shifted window attention to capture long-range dependencies in GADF images, such as the periodic outer-ring fault features associated with the diagonal elements of the Gram matrix. The CNN-GAM branch enhances local transient textures through convolutional spatial attention, such as the impulse signals of roller faults in Dataset B, and suppresses noise-dominated frequency bands through channel attention. After the two features are fused by channel concatenation, the global structure perception and local discriminative cues work together, making the complete proposed model achieve significantly better accuracy on Datasets A and B than single-module combinations. Compared with the pure attention model (Swin Transformer-GAM), the B-spline non-linear mapping of KAN-Swin Transformer compensates for the deficiency of MLP in modeling high-frequency features, improving the accuracy by 1.14%. Compared with the single-branch structures of KAN-Swin Transformer or CNN-GAM, the double-branch fusion improves the accuracy by 6.57% - 8.57% under complex working conditions, verifying the necessity of global-local feature complementarity. Compared with traditional CNN-Transformer models, the proposed model can better accurately capture high-energy fault regions and suppress channel noise frequency bands, thus achieving better accuracy.

[0080] Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art and related fields without creative efforts shall fall within the protection scope of the present invention. The structures, devices, and operation methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art without special instructions and limitations.

Claims

1. An aerospace bearing fault diagnosis method based on a GADF-driven KAN-Swin Transformer dual-branch network, characterized in that, It includes the following steps: Step 1: Signal acquisition and GADF image generation. The bearing vibration signals in different health states are acquired through channel sensors. The original time-series data is windowed to construct a training / test dataset. The Gramian Angular Difference Field (GADF) method is used to convert each signal window into a two-dimensional image, and then polar coordinate transformation is performed to retain its time-series correlation features. The generated GADF images are normalized and stored as an RGB format dataset, and the training set, validation set, and test set are divided according to the ratio of 7:2:

1. Step 2: Hybrid feature extraction and model training. The training set of GADF images is input into the KANSwinT-GAM architecture for joint feature learning, and the model processes through a dual-branch parallel manner. Step 3: Fault diagnosis and verification. The test set of GADF images is input into the trained model for health state recognition. The KAN-Swin Transformer and GAM framework achieve collaborative extraction of fault features through a multi-level feature fusion mechanism, combining the structure-aware ability of the KAN-Swin Transformer and the texture-sensitive characteristics of the CNN-GAM branch.

2. The aerospace bearing fault diagnosis method based on the GADF-driven KAN-Swin Transformer dual-branch network according to claim 1, wherein: The size of the GADF images in Step 1 is 224×224.

3. The aerospace bearing fault diagnosis method based on the GADF-driven KAN-Swin Transformer dual-branch network according to claim 1, wherein: The dual-branch parallel processing in Step 2 is as follows: KAN-Swin Transformer branch: Capturing local spatial dependencies based on the shifted window self-attention mechanism, processing 4×4 non-overlapping image patches using a hierarchical structure, using the KAN network instead of the traditional MLP for classification, and optimizing the non-linear mapping of features through a dynamic B-spline grid to generate multi-scale features. CNN-GAM branch: Extracting global texture features through a CNN backbone network that stacks Conv2D, ReLU, and GAM attention layers, where the GAM mechanism enhances discriminative features through the concatenation of channel attention and spatial attention. The hierarchical features output by the KAN-Swin Transformer branch and the high-dimensional global features of the CNN-GAM branch are cross-modally fused through channel concatenation and adaptive average pooling.

Images