Bearing fault noise immunity diagnosis method based on IPDSCS and Swin Transformer

Abstract

The invention discloses a bearing fault noise immunity diagnosis method based on IPDSCS and Swin Transformer, and the method comprises the steps: collecting vibration signals of a bearing under different working conditions, and obtaining a plurality of groups of original vibration signal data; performing wavelet transform on the original vibration signal data, and converting the original vibration signal data into a time-frequency image; constructing a feature extraction model based on an improved inverted pyramid depth separable convolution sequence, and extracting the time-frequency image to obtain a multi-dimensional feature vector; using the feature vector as input data, performing model training based on a Swin Transform deep learning network, and establishing a bearing fault diagnosis model; and inputting a to-be-diagnosed bearing vibration signal into the bearing fault diagnosis model, and outputting a corresponding fault category. According to the method, the noise immunity of signal processing is improved through wavelet transform denoising and an IPDSCS feature extraction method; a Swin Transform network is adopted to carry out multi-scale modeling on the multi-dimensional features, so that the accuracy and robustness of fault diagnosis are remarkably improved; and high-precision diagnosis of bearing faults is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and more specifically to a bearing fault noise-resistant diagnosis method based on inverted pyramid depthwise separable convolution sequence (IPDSCS) and Swin Transformer. Background Art

[0002] With the rapid development of industrial automation and digitalization, rolling bearings, as core components of mechanical systems, play an important role in industrial production. However, long-term operation and high-load operation of bearings can easily lead to various faults, affecting equipment performance and safety. Traditional bearing fault diagnosis methods mainly rely on manual experience and professional knowledge, and have problems such as low efficiency and poor accuracy. In recent years, deep learning technology has been widely used in the field of bearing fault diagnosis, which can effectively extract fault features and improve diagnostic accuracy and efficiency.

[0003] However, in actual industrial applications, there is inevitably a lot of noise in the collected signals, which will mask the rolling bearing fault signals and make fault diagnosis more complicated. Therefore, how to overcome the interference of noise, better process these complex data, fully mine the information in the data, and provide a more comprehensive perspective for fault diagnosis is an urgent problem that technicians in this field need to solve. Summary of the invention

[0004] In view of this, the present invention provides a bearing fault noise-resistant diagnosis method based on IPDSCS and Swin Transformer, which can solve the above problems. The method can effectively extract key features in vibration signals and realize accurate fault diagnosis in a noisy environment.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] The embodiment of the present invention provides a bearing fault noise immunity diagnosis method based on IPDSCS and Swin Transformer, comprising the following steps:

[0007] S1. Using an acceleration sensor to collect vibration signals of the bearing under different working conditions to obtain multiple groups of original vibration signal data;

[0008] S2, performing wavelet transform on the original vibration signal data to convert it into a time-frequency image;

[0009] S3, constructing a feature extraction model based on an improved inverted pyramid depth-separable convolution sequence, extracting the time-frequency image, and obtaining a multi-dimensional feature vector;

[0010] S4, using the feature vector as input data, performing model training based on the Swin Transformer deep learning network to establish a bearing fault diagnosis model that can adapt to different noise levels;

[0011] S5. Input the bearing vibration signal to be diagnosed into the bearing fault diagnosis model, and output the corresponding fault category.

[0012] It can be seen from the above technical solutions that compared with the prior art, the present invention has the following technical advantages:

[0013] (1) A bearing fault diagnosis method combining inverted pyramid deep separable convolution sequence (IPDSCS) and Swin Transformer is proposed. This method performs multi-scale analysis on one-dimensional vibration signals through wavelet transform, and then uses deep learning technology to extract key features in the signal, thus achieving high-precision diagnosis of bearing faults.

[0014] (2) Aiming at the common noise interference problem in industrial applications, the method of the present invention can effectively extract useful fault features from noise. Through the design of grouped convolution and dynamically separable blocks, the model improves the ability to extract bearing fault feature information while maintaining its lightweight, and enhances the robustness of the model in complex noise environments.

[0015] (3) The method of the present invention is superior to existing methods in terms of bearing fault diagnosis accuracy, robustness, and generalization ability, and has great potential and value in practical industrial applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0017] Figure 1 A flow chart of the bearing fault noise immunity diagnosis method based on IPDSCS and Swin Transformer provided by the present invention.

[0018] Figure 2 A diagram of the depthwise separable convolutional structure provided by the present invention.

[0019] Figure 3 This is a diagram of the residual depth separable block structure provided by the present invention.

[0020] Figure 4 A diagram of a dynamically separable block structure provided by the present invention.

[0021] Figure 5 This is a structural diagram of the Swin-Transformer provided by the present invention.

[0022] Figure 6 This is a structural diagram of the bearing fault detection model based on IPDSCS and Swin Transformer provided by the present invention.

[0023] Figure 7 Schematic diagram of the grouped convolution provided by the present invention for performing RGB three-channel learning on the fault TF image.

[0024] Figure 8 This is a diagnostic flow chart of the IPDSCS-SwinT method provided by the present invention.

[0025] Fig. 9 Schematic diagram of the test bench of the CWRU data set in Example 1.

[0026] Fig.10 This is the class activation mapping heat map of the proposed method IPDSCS-SwinT provided by the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The embodiment of the present invention discloses a bearing fault noise immunity diagnosis method based on IPDSCS and Swin Transformer, referring to Figure 1 As shown, the following steps S1 to S5 are included:

[0029] S1. Using an acceleration sensor to collect vibration signals of the bearing under different working conditions to obtain multiple groups of original vibration signal data;

[0030] In this step, the acceleration sensor can be used to collect vibration signals of a certain type of bearing under three different working conditions (normal, inner ring fault, outer ring fault, and rolling element fault). The sampling frequency is 25.6kHz, and the length of each data set is 1024 points.

[0031] S2, performing wavelet transform on the original vibration signal data to convert it into a time-frequency image;

[0032] Among them, wavelet transform is a powerful signal analysis tool that combines time and frequency analysis. It is particularly suitable for processing non-stationary signals and capturing local features in signals. The core advantage of this analysis method lies in its multi-resolution capability, that is, it can provide detailed information of the signal at different scales at the same time. Wavelet transform can adaptively focus on different parts of the signal by changing the size and shape of the window instead of keeping it fixed. Under the adjustment of time window and frequency window, wavelet transform can achieve fine time resolution in high-frequency area and provide good frequency resolution in low-frequency area. Through the scaling and translation operations of wavelet transform, the signal is gradually refined at multiple scales, thereby achieving the focus on any details of the signal. The sum of the product of the wavelet transform signal f(t) and the scaled and translated wavelet function (ψ(·)) over the entire period of the signal is shown in formula (1).

[0033]

[0034] Where a is the scale of the scale, τ is the translation position, ψ(·) is the wavelet function, and t is the time. The sum of all signal values ​​in the time domain multiplied and added by the wavelet function is the corresponding wavelet coefficient.

[0035] Among the many wavelet basis functions, the cmor3-3 wavelet stands out due to its unique properties and exhibits small differences, which makes it widely used in the field of signal processing. The cmor wavelet, as a Gaussian wave modulated by a single-frequency complex sine, not only performs well in the time domain, but also has excellent characteristics in the frequency domain. This wavelet can provide high-resolution time-frequency analysis when processing signals, and is particularly suitable for capturing and analyzing local features of signals.

[0036] In this embodiment, cmor3-3 wavelet is selected as the analysis tool. This choice is based on the excellent ability of cmor3-3 wavelet in maintaining the high resolution of signal features. It can accurately focus on the details of the signal, thereby improving the accuracy of time-frequency analysis. By applying cmor3-3 wavelet, useful information can be extracted from noise more effectively, providing strong support for intelligent fault diagnosis of rolling bearings. Its expression is shown in formula (2).

[0037]

[0038] Among them, F c is the center frequency of the wavelet, F b is the bandwidth parameter. i is a fictitious unit. x represents the original vibration signal data. c =F b =3.

[0039] S3, constructing a feature extraction model based on an improved inverted pyramid depth-separable convolution sequence, extracting the time-frequency image, and obtaining a multi-dimensional feature vector;

[0040] Among them, based on the improved inverted pyramid depth separable convolution sequence, including:

[0041] Grouped convolution layer, used to perform grouped convolution operations on time-frequency images;

[0042] Residual depthwise separable block for feature extraction on the output of grouped convolutional layers;

[0043] Dynamic separable block, used to perform feature extraction on the output of the residual depth separable block.

[0044] 1). Grouped convolution is an efficient convolution operation technique in convolutional neural networks (CNNs). It divides the channels of the input data into multiple groups, each of which is equipped with an independent set of convolution kernels. In this setting, the convolution kernels within each group are shared within the entire group, that is, multiple channels within the same group will use the same convolution kernel for data processing. This approach allows the network to reduce the amount of computation and the number of parameters while maintaining performance. Through grouped convolution, specialized convolution kernels can be applied to different data channels, which enhances the network's adaptability to input features. In addition, since the convolution kernels within each group are shared, the storage requirements of the model can be reduced and computational efficiency can be improved. This technique is particularly suitable for processing high-dimensional data with a large number of channels, and can reduce the complexity of the model while maintaining network performance.

[0045] The size of ordinary convolution is WH and C channels. The input of grouped convolution can be divided into n groups of feature map size: Each convolution kernel is also divided into n groups: The output feature map size is W′×H′×g. Therefore, the number of parameters and the amount of computation of grouped convolution are shown in formulas (3) and (4).

[0046]

[0047]

[0048] In the formula, params represents the number of parameters of grouped convolution; FLOPs represents the amount of computation of grouped convolution; k represents the number of convolution kernels in each group; n represents the number of groups; C represents the number of input channels; W', H' represents the width and height of the output feature map.

[0049] The grouped convolution technique significantly reduces the number of model parameters and computational burden by distributing convolution operations into multiple groups. This method is applied to wavelet transformed image data, which usually exists in the form of three RGB color channels. By treating these three channels as independent groups for grouped convolution, not only the number of parameters required by the network in the feature extraction stage is reduced, but also the precision of the extracted fault information is improved.

[0050] In addition, a global depthwise convolution (GDC) is introduced in each channel. GDC is a special convolution operation in which the number of groups G is equal to the number of channels N and C of the input and output, i.e., G=N=C. This means that the convolution kernel within each channel group will operate independently of other channel groups, and the size of the convolution kernel matches the size of the input feature map, i.e., K=H=W. This design allows GDC to assign a learnable weight to each position in the signal, thereby enabling adaptive and targeted learning of different areas of the vibration signal.

[0051] 2). Residual depth separable block

[0052] Depthwise Separable Convolution is an efficient convolution technique used in convolutional neural networks (CNNs). It reduces the number of model parameters and computational burden by simplifying convolution operations, thereby improving the model's operating efficiency and lightweight level. Figure 2 As shown. The first is the depthwise convolution, which applies a dedicated convolution kernel (Filters) to each input channel (Channel Input) independently, which means that the convolution operation of each channel is performed separately and does not involve information integration between different channels. This method allows the convolution operation to be performed only in the dimension of a single channel, thereby significantly reducing the number of parameters and the amount of calculation. After that, the pointwise convolution uses a 1x1 convolution kernel. Its function is to integrate the features of different channels based on the output of the depthwise convolution to achieve cross-channel information interaction. In this way, the pointwise convolution combines the information of each channel to form a more comprehensive image feature representation (Maps). Since the depthwise convolution and pointwise convolution can be performed independently, the depthwise separable convolution greatly reduces the number of parameters required. Compared with the traditional convolution operation, this not only reduces the computational burden of the model, but also makes the model more lightweight.

[0053] This embodiment proposes a new residual depth separable module, such as Figure 3As shown. The residual network (ResidualNetwork, often referred to as ResNet) is a deep neural network architecture. It can be easily expanded to very deep layers without performance degradation. This design makes ResNet very effective in processing large-scale image datasets and performing complex tasks. The residual deep separable module integrates the deep separable convolution technology into the residual block and extracts features in the jump connection through 1x1 separable convolution. This not only enhances the back propagation of the gradient, but also ensures the differentiation between input and output, thereby improving the expressiveness of the network. In addition, in order to further improve the performance of the module, layer normalization (LayerNormalization, LN) and activation function GeLU are used to construct the residual block. LN normalizes each feature of each sample, making the model insensitive to changes in the distribution of the input data, making it more stable when processing data with different distributions. The layer normalization operation is independent of the batch size, so it is particularly suitable for variable-length sequence data and can effectively extract the specific features of fault samples. By counting all neuron inputs in each layer on each training sample and normalizing the inputs with the obtained mean and variance, it helps to improve the generalization ability and robustness of the model.

[0054] The mean is shown in formula (5), and the variance is shown in formula (6).

[0055]

[0056]

[0057] H is the number of hidden neurons in the layer, where is the normalized sum of the inputs to the ith hidden unit in layer l. Under layer normalization, all hidden neurons in the same layer have the same normalization terms μ and σ.

[0058] The activation function GeLU is an activation function that relies on the probability statistics of the input itself to provide random regularity for the activation function while maintaining the input information. GeLU is shown in formula (7).

[0059] GeLU(x)=x*P(X,,x)=x*Φ(x) (7)

[0060] Where Φ(x) refers to the cumulative function of the Gaussian normal distribution of x. Furthermore, the function can be obtained as shown in formula (9).

[0061]

[0062] Where μ and σ represent the mean and standard deviation of the normal distribution respectively. Since the above function cannot be calculated directly, the GeLU function can be approximately expressed as formula (10).

[0063]

[0064] x represents the input of the activation function GeLU.

[0065] This means that the activation function GeLU is a smooth and differentiable activation function. Its derivative is available over the entire real number range and will not saturate over the entire real number range, which means that its derivative will not approach zero. This property is crucial for neural network training using optimization algorithms such as gradient descent. Because the non-saturated activation function can ensure the stable propagation of the gradient, thereby improving the efficiency and stability of network training. In contrast, although ReLU performs well for positive inputs, its derivative is zero for negative inputs. This may cause the gradient to disappear during backpropagation and affect the learning efficiency of the network. In addition, ReLU is a piecewise linear function, which is not smooth in processing positive and negative inputs, which may introduce instability during the optimization process. The smooth nature of GeLU helps alleviate these problems, making the optimization process smoother and improving the network's ability to learn complex function mappings.

[0066] 3). Dynamically separable blocks

[0067] Dynamic Convolution is a variant of Convolutional Neural Network (CNN). It introduces adjustable dynamic weights to make the convolution kernel perform adaptive parameter adjustments based on the characteristics of the input data. Compared with the traditional fixed-weight convolution kernel, this dynamic adjustment mechanism gives the network greater adaptability and enables it to respond more flexibly to different input scenarios.

[0068] In order to construct a model that is both lightweight and can effectively extract bearing fault characteristics, this embodiment designs a novel dynamically separable block. Figure 4 As shown in Figure 1, it relies on the attention mechanism to dynamically aggregate multiple convolution kernels into a new separable convolution weight matrix based on the input data. As shown in formula (11), this process essentially increases the number of channels, expands it by K times, and applies dynamic attention to K channel groups.

[0069] y=g(W T (x)x+b(x)) (10)

[0070]

[0071] In the formula, y represents a static perceptron, which is used to learn the importance weights of input features; x represents an input feature map, which represents the original signal or the features after group convolution; W represents a convolution kernel set, which contains K different convolution kernels. b represents a bias term, which is used to adjust the output of the convolution kernel; π represents an attention weight, which is used to control the degree of fusion of different convolution kernels; k is an index, which represents different convolution kernels.

[0072] Define y as a static perceptron, K convolution kernels {W k} and attention weight {π k (x)} linear aggregation. The dynamic separable module is introduced to improve the adaptability of the network to the input data without increasing the depth or width of the network. This module works through parallel separable convolution kernels that can share output channels, thereby avoiding the increase in the number of network parameters. The purpose is to implement an adaptive receptive field adjustment mechanism. When processing the input data after two-dimensional wavelet transformation, the module can dynamically adjust its receptive field according to the different position and scale characteristics in the data. This adaptive adjustment allows the network to more effectively capture and understand the structural information of the input data, thereby improving the ability to extract data features. In addition, since the design of the dynamic separable block focuses on working without expanding the channel depth, it helps to control the number of parameters of the network, thereby reducing the risk of overfitting. This design not only enables the convolution kernel to adapt to the changing input characteristics more flexibly, but also enhances the adaptability and generalization ability of the entire network.

[0073] S4, using the feature vector as input data, performing model training based on the Swin Transformer deep learning network to establish a bearing fault diagnosis model that can adapt to different noise levels;

[0074] In this step, Swin-Transformer is a deep neural network model based on the attention mechanism. It relies on the attention mechanism and can efficiently process large image data. Figure 5 As shown, including:

[0075] Windowed multi-head self-attention mechanism layer to capture long-range dependencies of input features;

[0076] The layer normalization layer is used to normalize the output of the windowed multi-head self-attention mechanism;

[0077] The MLP layer is used to extract features from the output of the layer normalization layer.

[0078] The core feature of this model is the windowed multi-head self-attention mechanism, which combines the advantages of local attention and multi-head attention to improve computational efficiency and storage efficiency. By using multiple parallel self-attention heads, SwinTransformer can learn data features from different perspectives at the same time. Each self-attention head focuses on a subset of the input data, namely a window. This design enables the model to capture local features while maintaining sensitivity to global context. In addition, the multi-head mechanism allows the model to learn in parallel in different representation subspaces, enhancing the understanding and representation of the input data. The windowed multi-head self-attention mechanism is a flexible neural network design that allows the model to balance attention to global and local information according to the needs of the processing task. Smaller windows can capture more local details, while larger windows can contain a wider range of data information, thereby better understanding global patterns. By adjusting the size of the window, the model can adapt to different data characteristics and task requirements. The attention mechanism is shown in formula (12).

[0079]

[0080] in, For query, key key and value value, M 2 is the total number of patches in a window, d is the dimension of each head, and is the relative position bias, which is used to characterize the relative position between patches; d k Dimensions representing queries, keys, and values.

[0081] In this embodiment, the extracted feature vector is input into the Swin Transformer network for model training. For example, the model includes 4 layers of local attention modules and 2 layers of global feature fusion modules, the learning rate is set to 0.001, and the training batch is 100. After the training is completed, the test data is input into the model, and the accuracy of the classification result reaches 98.7%.

[0082] S5. Input the bearing vibration signal to be diagnosed into the bearing fault diagnosis model, and output the corresponding fault category.

[0083] In this method, first, the original vibration signal is converted into a time-frequency image using wavelet transform, and the signal is converted from the time domain to the time-frequency domain, providing rich information for subsequent feature extraction. Subsequently, based on the improved deep separable convolution module, the improved module is specifically used to extract bearing fault features in a noisy environment. With this design, the model can more effectively separate useful fault features from noise. In addition, the introduction of Swin Transformer further enhances the model's ability to capture long-distance dependencies. Finally, the bearing fault diagnosis task in a noisy environment is carried out and trained and optimized accordingly.

[0084] This method combines the Inverted Pyramid Depthwise Separable Convolutional Sequence (IPDSCS) with the Swin Transformer to form an efficient fault diagnosis model. The core of the model is two innovative network structures: residual depthwise separable blocks and dynamic separable blocks. The introduction of these structures aims to improve the performance of convolutional neural networks in feature extraction while keeping the model parameters and complexity controllable.

[0085] The residual depthwise separable block leverages the concept of depthwise separable convolutions to reduce computational burden and parameter size, thereby enabling deeper feature extraction without significantly increasing model complexity. The module also employs residual connections to facilitate the training of deeper networks by alleviating the vanishing gradient problem, thereby building more accurate diagnostic models.

[0086] On the other hand, the dynamic separable block introduces an adaptive mechanism that adjusts the convolutional filters according to the input expert number, making the network focus more on informative features and less on noise. This dynamic adjustment provides a degree of flexibility to the model, which is critical for handling the variations and noise present in bearing vibration signals.

[0087] In addition, the introduction of Swin Transformer provides the model with powerful processing capabilities, especially its windowed multi-head self-attention mechanism, which enables the model to process large-size image data more finely and capture the global and local features in image or sequence data, thereby improving the accuracy and robustness of fault diagnosis.

[0088] Reference Figure 6As shown in Figure 1, this method combines the inverted pyramid deep separable convolution sequence with the Swin-Transformer model in series, and the parameter design is shown in Table 1. The core design goal of this method is to solve the limitations of traditional bearing fault diagnosis methods in feature extraction and model building. Through this novel combination method, bearing faults can be diagnosed more accurately and efficiently. IPDSCS can capture detailed information of input data at different scales and realize multi-scale feature extraction through its inverted pyramid structure and deep separable convolution technology. The Swin Transformer model uses its attention-based design to effectively process sequence data and capture long-distance dependencies in the data. The serial combination of these two technologies not only enhances the model's ability to capture complex features in bearing vibration signals, but also improves the model's robustness to noise and other interference factors. The advantage of this integrated method is that it can make full use of the advantages of both to achieve deeper feature learning and more accurate fault pattern recognition.

[0089] First, the inverted pyramid depthwise separable convolution sequence is a new type of feature extraction module, which combines the inverted pyramid structure with the depthwise separable convolution technique. The inverted pyramid structure can capture different details and features of the input data at different levels, thereby achieving multi-scale feature extraction. Figure 7 As shown in the figure, the grouped convolution at different scales groups the wavelet transformed image for the RGB channels. The three groups of RGB channels represent the features of the image in the red, green and blue channels respectively. Grouped convolution learns features independently on each channel. By performing convolution operations on each group of channels separately, the network can better capture the feature changes of the image in different color channels, thereby more accurately understanding the image content. The convolution operation of each channel is performed independently, so the features of different channels can be processed in parallel, which speeds up the training and reasoning of the model. Grouped convolution is conducive to better utilizing image features, improving the expressiveness of the model, reducing the computational complexity and the number of parameters, and sending the features of different channels to subsequent different depth-wise separable convolution blocks. Depth-wise separable convolution decomposes the convolution operation into depth-wise convolution and point-wise convolution, effectively reducing the number of parameters and computational complexity of the model. Residual connection allows the model to directly learn the residual between input and output, which helps to alleviate the gradient vanishing problem and accelerate the training process.

[0090] Then, the dynamic separable block can adaptively and dynamically adjust the number of convolution kernels, making the model more flexible and adaptable. The dynamic separable block introduces a dynamic convolution mechanism, and the default setting of the number of dynamic channels is 3, so that the weight of the convolution kernel can be adaptively adjusted according to different parts of the input data. The dynamic separable block can adaptively adjust the weight of the convolution kernel according to the input data, so that the model can pay more attention to important features and reduce the sensitivity to irrelevant information. The risk of overfitting of the model is reduced and the generalization ability of the model is improved. Therefore, the inverted pyramid deep separable convolution sequence can extract rich frequency domain and time domain features, and extract more comprehensive feature information for the noise resistance diagnosis of bearing faults. Secondly, Swin-Transformer, as a neural network structure based on the self-attention mechanism, is particularly suitable for processing large-size images and sequence data. Its windowed local attention mechanism and hierarchical design can effectively capture long-distance dependencies in images or sequence data, improving the modeling ability and generalization performance of the model.

[0091] IPDSCS-Swin Trans can make full use of the advantages of both to achieve stronger feature extraction and model building, thereby further improving the accuracy and stability of bearing fault diagnosis and realizing end-to-end fault diagnosis.

[0092] Table 1 Structural parameter design of IPDSCS-SwinT method

[0093]

[0094]

[0095] In summary, the present invention proposes a bearing fault diagnosis method combining inverted pyramid depthwise separable convolution sequence (IPDSCS) and SwinTransformer. The process framework based on the IPDSCS-SwinT method is as follows: Figure 8 As shown, the diagnostic procedure is outlined below.

[0096] 1. Through end-to-end one-dimensional data collection, complete information can be obtained directly from the data source, which ensures the integrity and consistency of the data and provides reliable support for data-based decision-making.

[0097] 2. Through cmor3-3 wavelet transformation, the original one-dimensional data can be analyzed in time and frequency and converted into TF images. This transformation analyzes the data in two dimensions, time and frequency, so that we can have a more comprehensive understanding of the characteristics and changing patterns of the data.

[0098] 3. When performing wavelet transform, a noisy TF image is generated by adding noise. The purpose of this is to simulate the noise environment that may exist in actual work, so as to more realistically reflect the data situation.

[0099] 4. Fault diagnosis and noise resistance Fault diagnosis is achieved by analyzing and processing TF image data. TF images are classified and identified to determine whether the data has anomalies or faults.

[0100] 5. Diagnostic results can be displayed through visualization and analysis. By presenting the diagnostic results in the form of images or charts, the status and fault conditions of the data can be analyzed intuitively.

[0101] Next, an example is used to verify the advantages of the IPDSCS-SwinT method proposed in the present invention on a benchmark bearing fault diagnosis dataset. The benchmark dataset is the Case Western Reserve University (CWRU) bearing dataset. The IPDSCS-SwinT method of the present invention is written using the TensorFlow 2.5.0 framework in Python 3.8. All experiments were conducted in the same operating environment, with computer configurations including Windows 1016GB RAM, i9-9900K CPU, and NVIDIA RTX 2070SUPER 8GB graphics processor.

[0102] Different methods of troubleshooting

[0103] To verify the superiority of the method proposed in this paper, we compared the following deep learning methods. To reasonably compare the characteristics of the methods, the parameters of all compared models are kept consistent with the original text as much as possible. The detailed description of the comparison method is as follows.

[0104] a.AlexNet: A deep convolutional neural network model proposed by Alex Krizhevsky et al. in 2012. It is a convolutional neural network with 8 learnable layers. It consists of multiple convolutional layers and pooling layers stacked together, and the Dropout regularization technology is introduced. The risk of overfitting is reduced by randomly setting the outputs of some neurons to zero. A local response normalization layer is used in the layer after activation to suppress the activity of neurons and enhance the robustness of the model. Moreover, AlexNet widely used ReLU (rectified linear unit) as the activation function for the first time. In the embodiment, GeLU is selected as the activation function for comparison.

[0105] b. InceptionV3: It is a deep convolutional neural network architecture proposed by the Google research team in 2015. InceptionV3 uses a structure called the Inception module, which is an efficient feature extraction module. The input is processed in parallel by convolution kernels of different scales and sizes, and then the results are spliced ​​together in the depth direction. Auxiliary classifiers are introduced in the middle layer to help the gradient propagate backward during the training process. And deep separable convolution is introduced in the network, which is a lightweight convolution operation. In the final stage, a global average pooling layer is used instead of a fully connected layer to reduce the number of parameters and reduce the risk of overfitting. InceptionV3 and the inverted pyramid structure in the method of the present invention are both a deformation of the Inception structure, and it is necessary to compare them.

[0106] c.VGG16: A deep convolutional neural network model proposed by the Visual Geometry Group of the University of Oxford in 2014. It consists of 16 layers of neural networks. VGG16 uses continuous convolutional layers to extract features. Compared with previous models, it has a deeper network structure. Each convolutional layer uses a small-sized 3x3 convolution kernel and a convolution operation with a stride of 1, which helps to increase the nonlinear expression ability of the network. VGG16 uses a maximum pooling layer to reduce the spatial dimension of the feature map and retain important feature information. Several fully connected layers are connected after the convolution layer to map high-level features to category labels. The network structure of VGG16 is relatively simple and clear, the convolution kernel size and stride are consistent, and the size of the feature map is halved after each pooling.

[0107] d.SwinTransformer: A new Transformer architecture proposed by Microsoft Research Asia in 2021. Swin Transformer introduces a cross-layer connection mechanism, which allows information exchange and transmission between different layers. The input image is divided into non-overlapping image blocks and a windowed processing method is used to avoid the high memory and computational complexity of traditional global attention calculations. Local attention calculations are performed between each window, and then integrated at the global level to achieve global perception of the entire image.

[0108] e.ShuffleNet: ShuffleNet is a lightweight deep neural network architecture proposed in 2018. It adopts an efficient network design to reduce the number of model parameters and computational cost through deep separable convolution and channel shuffle operations. ShuffleNet uses deep separable convolution operations to decompose the standard convolution into two steps: deep convolution and point-by-point convolution. It also introduces the channel shuffle operation to shuffle the feature maps of different channels, thereby increasing the communication and information transfer between features.

[0109] In the comparative experiment, we used the Adam optimization algorithm to train the model, which can automatically adjust the learning rate to adapt to the data parameters. We set 128 batches in the training, each batch contains 16 samples, and set the learning rate to 0.0001, which is a relatively small value that helps the model to converge stably during the training process. In order to ensure the reliability of the experimental results and reduce the impact of chance, we implemented a 10-fold cross-validation technique. In addition, we monitor the changes in the loss value during the training process. When the loss value tends to stabilize, that is, the performance of the model is no longer significantly improved, we choose to stop the iteration to avoid overfitting.

[0110] Example: CWRU bearing dataset

[0111] We selected the CWRU bearing dataset as one of the benchmark datasets. This dataset is a failure experiment of SKF6205-2RS rolling bearings. Fig. 9As shown. The entire test bench consists of fan side (Fan end) and drive side rolling bearings (Testbearing), dynamometer (Testbase), torque sensor (Torque Drive end transducer) and motor (Electric motor). We selected the drive side rolling bearing data as the experimental data. At a sampling frequency of 12kHz, four different motor loads were used to drive the faulty bearing, and the vibration acceleration signal of the bearing was collected by the acceleration sensor. Mainly through artificial electrical discharge machining, single-point groove faults with diameters of 0.007 inches, 0.014 inches, 0.021 inches and 0.028 inches were formed at the three positions of the ball (BA), inner ring (IR) and outer ring (OR) of the bearing. In the experiment, we selected fault data under 1hp, 1772r / min and 2hp, 1750r / min working conditions to set up three data sets, each of which contains 4 ball faults, 4 inner ring faults and 3 outer ring faults. The classification of data set samples is shown in Table 2. We use random sliding window sampling with a sliding window of 1024, that is, each sample has 1024 data points. Sliding window sampling can retain the timing information of time series data. Because it slides on the data at a fixed step size, it can obtain rich time domain features, frequency domain features, and time-frequency features. Each sampling considers the relationship between the previous and next data points, and there is no additional data processing cost. The ratio of training samples to test samples in each data set is 7:3.

[0112] Table 2 Classification of samples of CWRU rolling bearing dataset

[0113] category Fault type Fault size 0 Rolling element 0.007 1 Rolling element 0.014 2 Rolling element 0.021 3 Rolling element 0.028 4 Inner Circle 0.007 5 Inner Circle 0.014 6 Inner Circle 0.021 7 Inner Circle 0.028 8 Outer ring 0.007 9 Outer ring 0.014 10 Outer ring 0.021

[0114] 1Performance under strong noise

[0115] In industrial production, the vibration signals generated by bearings during operation are often disturbed and affected by environmental noise. This noise may mask fault information and make fault diagnosis difficult. Therefore, it is necessary to evaluate the performance of the proposed fault diagnosis method in a high-noise environment. We add Gaussian noise to the original signal to construct a composite signal with different signal-to-noise ratios (SNR). The definition of the signal-to-noise ratio is as follows:

[0116]

[0117] Where P signal is the original signal power, P noise is the noise power. From the formula, we can know that SNR and P noise Therefore, the more serious the signal is contaminated by noise, the smaller the SNR value is.

[0118] We selected a data set with a speed of 1750r / min and a load of 2hp for the experiment, and added 6, 3, 0, -3 and -6dB noise to the data set to simulate industrial operation. First, if the number of noise samples is sufficient, we can increase the signal-to-noise ratio in both training and test samples. The purpose of this is to allow the model to better adapt to high-noise environments during training, and to accurately identify fault signals and resist noise interference during the test phase. Secondly, for samples collected under noisy conditions, we only add the signal-to-noise ratio to the test samples. This is to simulate the situation in the actual working environment and ensure that the model can effectively operate and diagnose faults in a real noisy environment. Finally, for the limited samples collected under normal conditions, we only add the signal-to-noise ratio to the training samples. This method helps the model learn the fault characteristics under the influence of noise during training and improves the robustness and generalization ability of the model.

[0119] We add signal-to-noise ratio to both training and test samples to simulate the situation where noise is prevalent in real working environments and evaluate the performance of the model in the face of prevalent noise. Table 3 shows the fault diagnosis rate of each comparison method under 5 signal-to-noise ratios. Under 300 samples with a 6dB signal-to-noise ratio, the single bearing fault of each comparison method can reach 100%. However, when the number of samples decreases, it can be seen that the fault diagnosis effect of IPDSCS-SwinT is significant. VGG16 still has a good fault diagnosis rate and can have a certain degree of noise resistance in the TF image of bearing faults. Under 3dB and 0dB conditions: IPDSCS-SwinT maintains the highest accuracy, 99.93% and 99.30% respectively, showing the best performance under medium noise levels. VGG16 and InceptionV3 also performed well, with an accuracy slightly lower than IPDSCS-SwinT, but still maintained at around 99%. Even when the signal-to-noise ratio is low and the noise influence is large, VGG16 can still maintain a very high accuracy. As the SNR is further reduced to -3dB and -6dB, IPDSCS-SwinT still performs best, with accuracy rates of 98.89% and 98.63%, respectively. Swin Transformer and VGG16 also perform relatively well under -3dB conditions, but as the SNR is further reduced, the accuracy drops significantly. Under -6dB conditions, the accuracy of Swin Transformer (93.13%) is lower than that of IPDSCS (95.74%), which indicates that IPDSCS may have better performance under extreme noise conditions. ShuffleNet performs well under high SNR conditions, but its accuracy drops rapidly under low SNR conditions, indicating that it is more sensitive to noise. VGG16 and InceptionV3 also perform well under medium SNR conditions, but their performance in high noise environments is not as good as IPDSCS-SwinT. Swin Transformer has a high accuracy under medium SNR conditions, but its performance drops rapidly under extremely low SNR conditions. AlexNet and ShuffleNet have high accuracy under high signal-to-noise ratio conditions, but their accuracy drops significantly when the noise is large. IPDSCS-SwinT combines the advantages of the inverted pyramid depth-separable convolution sequence and Swin Transformer, which can effectively extract key features in vibration signals and achieve accurate fault diagnosis in the presence of noise. IPDSCS-SwinT shows excellent performance under various signal-to-noise ratio conditions, especially the stability and accuracy under low signal-to-noise ratio conditions, highlighting the excellent robustness and generalization ability of this method in different noise environments.

[0120] Table 3 Single noise fault diagnosis rate

[0121]

[0122] Table 4 shows the fault diagnosis accuracy of 200 samples of different deep learning methods under different signal-to-noise ratio conditions without adding noise to the test samples. At 6dB, the accuracy of AlexNet is 99.16%, and as the signal-to-noise ratio decreases, the accuracy gradually decreases, and drops to 86.67% at -6dB. This shows that AlexNet's resistance to noise decreases with the decrease of the signal-to-noise ratio. VGG16 has the highest accuracy at 6dB, reaching 99.29%, and still maintains a high accuracy of 89.36% when the signal-to-noise ratio drops to -6dB. This shows that VGG16 has good stability under different noise levels. InceptionV3 performs similarly to VGG16 under high signal-to-noise ratio, but its accuracy drops faster under low signal-to-noise ratio conditions. It drops from 99.13% at 6dB to 85.56% at -6dB, indicating that it is more sensitive to noise than VGG16. ShuffleNet has an accuracy of 99.14% at 6dB, and as the SNR decreases, the accuracy drops to 87.23% at -6dB. Swin Transformer has an accuracy of 98.56% at 6dB, and drops to 83.54% at -6dB. Swin Transformer performs well under high SNR conditions, but its performance drops significantly under low SNR. IPDSCS has an accuracy of 98.08% at 6dB SNR and 84.24% at -6dB. The performance of IPDSCS remains relatively stable in the medium to low SNR range. IPDSCS-SwinT shows the best fault diagnosis accuracy under all tested SNR conditions, from 99.62% at 6dB to 90.73% at -6dB. Even under low SNR conditions, the accuracy of IPDSCS-SwinT drops less, showing the best robustness.

[0123] Table 4 Fault diagnosis rate of test samples without adding noise (200 samples)

[0124] Algorithms 6dB 3dB 0dB -3dB -6dB AlexNet 99.16±0.68 97.87±1.64 95.20±3.57 90.87±2.59 86.67±2.68 InceptionV3 99.13±0.52 97.86±1.48 95.88±3.59 91.49±3.63 85.56±2.99 VGG16 99.29±0.38 98.03±1.41 96.71±2.15 92.96±2.88 89.36±2.04 ShuffleNet 99.14±0.55 97.25±1.61 95.51±2.59 91.03±2.94 87.23±2.98 SwinTransformer 98.56±0.81 96.66±2.27 94.03±3.79 89.36±2.83 83.54±2.96 IPDSCS 98.08±0.87 95.39±3.86 93.62±4.59 88.04±3.91 84.24±3.83 IPDSCS-SwinT 99.62±0.29 98.43±1.42 96.99±2.19 93.54±2.99 90.73±3.11

[0125] Table 5 provides the fault diagnosis accuracy of different deep learning methods under different signal-to-noise ratio conditions with 60 samples without adding noise to the training samples. The diagnostic accuracy of AlexNet at 6dB is 96.71%, and the diagnostic accuracy at -6dB is 83.70%. This shows that the greater the noise interference of AlexNet, the more severe the decline in diagnostic performance. VGG16 shows good robustness, and the accuracy decreases less under different signal-to-noise ratio conditions. AlexNet and InceptionV3 perform well under high signal-to-noise ratio conditions, but the accuracy decreases rapidly under low signal-to-noise ratio conditions, indicating that the resistance to noise is weaker when the number of samples is small. ShuffleNet and Swin Transformer have high accuracy under high signal-to-noise ratio conditions, but the performance decreases more significantly under extremely low signal-to-noise ratios, showing sensitivity to noise. IPDSCS shows relatively stable performance, but is not as good as IPDSCS-SwinT under low signal-to-noise ratio conditions.

[0126] IPDSCS-SwinT shows the highest fault diagnosis accuracy under all SNR conditions, from 98.07% at 6dB to 86.29% at -6dB. Even under the conditions of small number of samples and low SNR, the accuracy of IPDSCS-SwinT decreases slightly. With a small number of samples, IPDSCS-SwinT still shows excellent performance under various SNR conditions, especially in low SNR environment, and its ability to maintain accuracy shows that the model has very strong robustness.

[0127] Table 5 Fault diagnosis rate of training samples without noise (60 samples)

[0128] Algorithms 6dB 3dB 0dB -3dB -6dB AlexNet 96.71±1.72 93.86±1.87 96.99±2.64 87.82±3.11 83.70±2.89 InceptionV3 96.38±1.46 93.62±2.47 89.18±2.98 87.76±3.11 83.33±3.99 VGG16 97.68±1.17 94.34±1.06 90.94±2.34 89.02±2.55 84.34±3.01 ShuffleNet 96.33±1.69 93.84±1.99 90.00±2.31 87.23±2.67 81.90±3.97 SwinTransformer 93.60±2.48 90.88±2.51 88.05±2.84 86.87±4.31 81.43±4.94 IPDSCS 94.56±2.01 90.99±2.65 88.17±3.14 87.01±3.54 82.34±4.31 IPDSCS-SwinT 98.07±0.99 84.86±1.24 91.28±2.98 89.93±2.99 86.29±3.84

[0129] We used Class Activation Mapping (CAM) heatmaps for visualization, showing the areas that IPDSCS-SwinT focuses on when making a diagnosis, such as Fig.10As shown. The ability to clearly highlight fault-related signal features from CAM and maintain this ability in a noisy environment would indicate its significant advantage in bearing fault diagnosis tasks. CAM can accurately point out the specific frequency band or time domain features where the fault occurs, indicating that the IPDSCS-SwinT method has the ability to accurately locate the type of fault. In the presence of noise and other interference, CAM is still able to highlight relevant fault features, which would prove that the IPDSCS-SwinT model is robust to noise and interference in practical applications. The heat map of IPDSCS-SwinT can consistently highlight signal features corresponding to known fault modes, indicating that the model is able to learn complex relationships related to bearing faults.

[0130] 2. Performance under complex working conditions

[0131] In actual industrial environments, the complex operation of bearings is often affected by a variety of factors, including sudden changes in speed and load and the presence of noise. In order to more comprehensively study the impact of these factors on bearing performance, we selected two different working conditions: a speed of 1750r / min and a load of 2hp, and a speed of 1772r / min and a load of 1hp for experiments. First, a sudden change in speed may cause changes in the bearing force and vibration frequency, thereby affecting the working state and performance of the bearing. Secondly, a sudden change in load is also a common situation in industrial environments, which may cause changes in the force borne by the bearing and the load distribution of the rotor, thereby affecting the operating state of the bearing. Finally, these complex situations are often accompanied by the presence of noise, which may come from multiple factors such as the operation of the bearing itself and the interference of the external environment. Therefore, in the experiment, we consider the impact of noise on bearing performance and take corresponding measures to reduce the impact of noise in order to more accurately evaluate the working state and performance of the bearing. Through these experiments, we can better understand the working characteristics of bearings under different working conditions and provide reliable bearing fault diagnosis and prediction for industrial production.

[0132] Evaluation indicators are used to measure the performance and effectiveness of the model, which can help determine the model's ability to detect and diagnose faults. Precision, recall, misclassification rate and F1 score are the main evaluation indicators.

[0133] Precision: Also known as the precision rate, it is an evaluation indicator for prediction results. It refers to the proportion of samples that are actually positive among the samples predicted by the model as positive.

[0134]

[0135] Recall: Also known as the recall rate, it is an evaluation indicator for the original sample.

[0136] It refers to the proportion of samples that are actually positive that are successfully predicted by the model as positive.

[0137]

[0138] False positive rate: Also known as false positive rate, it refers to the ratio of the number of samples with classification errors to the total number of samples. It is usually used to measure the error rate of the classification model. The smaller the value, the better the performance.

[0139]

[0140] F1 Score: It is the harmonic mean of precision and recall. It takes into account both precision and recall and is very useful for model evaluation in imbalanced datasets.

[0141]

[0142] In the above formula, TP (True Positive) means that an instance is a positive class and is predicted as a positive class; FN (False Negative) means that an instance is a positive class but is predicted as a negative class; FP (False Positive) means that an instance is a negative class but is predicted as a positive class; TN (True Negative) means that an instance is a negative class and is predicted as a negative class. P represents precision and R represents recall.

[0143] Table 6 reflects the diagnostic performance of each method under different signal-to-noise ratio (SNR) conditions using evaluation indicators when the number of samples is relatively small. AlexNet and VGG16 perform similarly in most evaluation indicators, but are slightly lower than IPDSCS-SwinT in F1 score. This indicates that their performance is slightly inferior when precision and recall are considered comprehensively. InceptionV3 performs the worst among all models, especially in false alarm rate, which may indicate that the model is more likely to produce wrong positive predictions in fault diagnosis. Swin Transformer and ShuffleNet also show relatively good performance, especially in precision and F1 score, which may mean that they have lower error rates and better balance in fault identification. IPDSCS-SwinT shows the best or near-best performance in all evaluation indicators. Especially in precision and F1 score, this shows that IPDSCS-SwinT has the highest accuracy and comprehensive performance in fault diagnosis tasks when the number of samples is 80.

[0144] Table 61772-1750 Variable load noise (variable working conditions) fault diagnosis evaluation indicators (80 samples)

[0145]

[0146]

[0147] When analyzing the CWRU bearing dataset under 1772-1750 variable load noise conditions, we used the t-SNE technique to visualize the fault feature representation of 200 samples at 0dB signal-to-noise ratio. These fault types were classified by 7 different deep learning methods. Among these methods, AlexNet, VGG16, and ShuffleNet showed mutual interference in their feature representation when distinguishing between 21-inch outer ring faults and 28-inch inner ring faults, which may be due to the fact that the signal features of these two fault types are too close in high-dimensional space, making it difficult for the models to distinguish. Similarly, all 7 methods also showed mutual interference in the feature information of spherical 14-inch faults and outer ring 14-inch faults, which further exacerbated the difficulty of accurate classification. However, under the same conditions, the IPDSCS-SwinT model showed significant advantages. The feature representation of the IPDSCS-SwinT model still maintains a good structure even at a low signal-to-noise ratio of 0dB, which means that the model can more effectively distinguish different fault types even under unfavorable conditions with large noise. The excellent performance of the IPDSCS-SwinT model in t-SNE visualization demonstrates its superior performance in the bearing variable load noise fault diagnosis task, especially its robustness and accuracy when dealing with low signal-to-noise ratio data.

[0148] 3. Performance under composite noise

[0149] In industrial environments, the operating status of bearings is often disturbed by various noises. These noises may come from different sources, including mechanical movement, electromagnetic interference, environmental vibration, etc. The simultaneous action of multiple noises makes the bearing signal more complex. In order to more accurately simulate the complex noise conditions in industrial environments, we selected a data set with a speed of 1772r / min and a load of 1hp for the experiment. By adding Gaussian white noise of different intensities, we can simulate the impact of noise from different sources in industrial environments on bearing signals, so as to more comprehensively evaluate the diagnostic status and performance of the model for bearing faults.

[0150] When diagnosing bearing faults under composite noise, we evaluated the performance of various comparison methods through confusion matrices. The results show that under variable noise and variable load conditions, the VGG16 model shows good performance. However, its diagnostic effect is reduced when dealing with complex situations of composite noise and multi-feature coupling. At the same time, the anti-interference performance of InceptionV3, VGG16, ShuffleNet and IPDSCS models under composite noise conditions is average. In contrast, although AlexNet's performance in Section 2 above is only at a medium level, it shows a stronger fault feature decoupling ability than VGG16 under composite noise. However, when AlexNet is compared with Swin Transformer and IPDSCS-SwinT, IPDSCS-SwinT has a more significant advantage by introducing an attention mechanism to more effectively learn multi-feature expressions under composite noise. Experiments verify that the attention mechanism of Swin Transformer provides significant performance improvement when dealing with multi-feature coupling problems. Therefore, the IPDSCS-SwinT model shows excellent ability in multi-feature decoupling of composite noise. It can not only accurately distinguish mutually interfering fault signals, but also show a low error rate in the confusion matrix, which shows that the model can effectively suppress noise and extract useful fault features. Even in a noisy environment, IPDSCS-SwinT can maintain good diagnostic capabilities. In addition, the confusion matrix of the IPDSCS-SwinT model under different noise conditions maintains good performance.

[0151] The present invention proposes a bearing fault diagnosis method based on inverted pyramid deep separable convolution sequence (IPDSCS) and SwinTransformer, which has achieved remarkable results in the accuracy and robustness of bearing fault diagnosis. Experiments on the CWRU bearing dataset show that the IPDSCS-SwinT method can maintain a high diagnostic accuracy even in low signal-to-noise ratio and composite noise environments. In addition, compared with other deep learning methods, IPDSCS-SwinT performs well in dealing with fault decoupling under multi-feature coupling and composite noise conditions. Moreover, when the number of samples is small, IPDSCS-SwinT can still maintain stable diagnostic performance, showing robustness to changes in the number of samples.

[0152] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A bearing fault noise immunity diagnosis method based on IPDSCS and Swin Transformer, characterized in that: The following steps are involved: S1. Using an acceleration sensor to collect vibration signals of the bearing under different working conditions to obtain multiple groups of original vibration signal data; S2, performing wavelet transform on the original vibration signal data to convert it into a time-frequency image; S3, constructing a feature extraction model based on an improved inverted pyramid depth-separable convolution sequence, extracting the time-frequency image, and obtaining a multi-dimensional feature vector; S4, using the feature vector as input data, performing model training based on the Swin Transformer deep learning network to establish a bearing fault diagnosis model that can adapt to different noise levels; S5. Input the bearing vibration signal to be diagnosed into the bearing fault diagnosis model, and output the corresponding fault category.

2. The method according to claim 1, characterized in that: In step S1, the multiple groups of original signal data include: Vibration signals corresponding to normal state, inner ring fault, outer ring fault and rolling element fault.

3. The method according to claim 1, characterized in that: In step S2, the wavelet transform is performed using cmor3-3 wavelet, and the formula is as follows: Among them, F c is the center frequency of the wavelet, F b is the bandwidth parameter; i is an imaginary unit; x represents the original vibration signal data.

4. The method according to claim 1, characterized in that: In the step S3, based on the improved inverted pyramid depth-separable convolution sequence, the method includes: Grouped convolution layer, used to perform grouped convolution operations on time-frequency images; Residual depthwise separable block for feature extraction on the output of grouped convolutional layers; Dynamic separable block, used to perform feature extraction on the output of the residual depth separable block.

5. The method according to claim 4, characterized in that: The grouped convolution layer performs grouped convolution operations on the RGB channels of the time-frequency image, including the following: The input of the grouped convolution each set of feature map sizes is divided into n groups: Each convolution kernel is also divided into n groups: The output feature map size is W′×H′×g; W, H represent the width and height of the input feature map; The number of parameters and the amount of computation for grouped convolution are shown in formulas (3) and (4): In the formula, params represents the number of parameters of grouped convolution; FLOPs represents the amount of computation of grouped convolution; k represents the number of convolution kernels in each group; n represents the number of groups; C represents the number of input channels; W', H' represents the width and height of the output feature map.

6. The method according to claim 4, characterized in that: The residual depth can be separated into blocks, including the following: Depth-wise separable convolution layer, used to perform depth-wise separable convolution operation on input features; Point-wise convolution layer, used to perform point-wise convolution operation on the output of the depth-wise separable convolution layer; Residual connections are used to connect the output of the depthwise separable convolutional layer to the output of the pointwise convolutional layer.

7. The method according to claim 6, characterized in that: The residual connection uses layer normalization and activation function GeLU; Layer normalization: By counting the inputs of all neurons in each layer on each training sample, the inputs are normalized using the mean and variance obtained. The mean is shown in formula (5) and the variance is shown in formula (6): Where H is the number of hidden neurons in the layer, is the normalized sum of inputs to the ith hidden unit in the lth layer; Under layer normalization, all hidden neurons in the same layer have the same normalization terms μ and σ; Activation function GeLU: is approximately expressed as formula (9): x represents the input of the activation function GeLU.

8. The method according to claim 4, characterized in that: The dynamically separable block includes the following: Dynamic convolution layer, used to dynamically adjust the convolution kernel weights according to the input features; An attention mechanism layer that weights the output of the dynamic convolutional layer; it scales it K times and applies dynamic attention to K channel groups; y=g(W T (x)x+b(x)) (10) In the formula, y represents a static perceptron, which is used to learn the importance weights of input features; x represents an input feature map, which represents the original signal or the features after group convolution; W represents a convolution kernel set, which contains K different convolution kernels. b represents a bias term, which is used to adjust the output of the convolution kernel; π represents an attention weight, which is used to control the degree of fusion of different convolution kernels; k is an index, representing different convolution kernels.

9. The method according to claim 1, characterized in that: In the step S4, based on the Swin Transformer deep learning network, the following is included: Windowed multi-head self-attention mechanism layer to capture long-range dependencies of input features; The layer normalization layer is used to normalize the output of the windowed multi-head self-attention mechanism; The MLP layer is used to extract features from the output of the layer normalization layer.

10. The method according to claim 9, characterized in that: The windowed multi-head self-attention mechanism layer adapts to different data characteristics and task requirements by adjusting the size of the window; the attention mechanism is shown in formula (12): in, is the information to be queried, relevant information of the information to be queried, and specific information related to the queried information; 2 is the total number of patches in a window, d k represents the dimensions of query, key, and value; b represents the relative position deviation, which is used to characterize the relative position between patches.

Images