A bearing fault anti-noise diagnosis method based on IPDSCS and Swin Transformer

Abstract

The application discloses a bearing fault noise resistance diagnosis method based on IPDSCS and a Swin Transformer, and comprises the following steps: collecting vibration signals of a bearing under different working conditions to obtain multiple groups of original vibration signal data; performing wavelet transform on the original vibration signal data to convert the original vibration signal data into time-frequency images; constructing an improved inverted-pyramid deep separable convolution sequence feature extraction model to extract the time-frequency images to obtain multi-dimensional feature vectors; taking the feature vectors as input data, performing model training based on a Swin Transformer deep learning network, and establishing a bearing fault diagnosis model; and inputting bearing vibration signals to be diagnosed into the bearing fault diagnosis model to output corresponding fault categories. The method improves the noise resistance of signal processing through wavelet transform denoising and an IPDSCS feature extraction method; multi-scale modeling is performed on multi-dimensional features by using a Swin Transformer network, and the precision and robustness of fault diagnosis are significantly improved; and high-precision diagnosis of bearing faults is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of bearing fault diagnosis, and more particularly to a bearing fault noise resistance diagnosis method based on an inverted pyramid deep separable convolution sequence (IPDSCS) and a Swin Transformer. BACKGROUND

[0002] With the rapid development of industrial automation and digitization, rolling bearings, as the core components of mechanical systems, play an important role in industrial production. However, long-term operation and high-load work of bearings can easily lead to various faults, affecting the performance and safety of equipment. 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 applied in the field of bearing fault diagnosis, which can effectively extract fault features and improve the accuracy and efficiency of diagnosis.

[0003] However, in actual industrial applications, a large amount of noise inevitably exists in the collected signals, which can mask the rolling bearing fault signals, making fault diagnosis more complex. Therefore, how to overcome the interference of noise, better process these complex data, and fully mine the information in the data to provide a more comprehensive perspective for fault diagnosis is a problem that needs to be solved by those skilled in the art. SUMMARY

[0004] Therefore, the application provides a bearing fault noise resistance diagnosis method based on IPDSCS and Swin Transformer, which can solve the above problems. The method can effectively extract key features in the vibration signal and realize accurate fault diagnosis in a noisy environment.

[0005] In order to achieve the above purpose, the application adopts the following technical solutions:

[0006] The application provides a bearing fault noise resistance diagnosis method based on IPDSCS and Swin Transformer, which includes the following steps:

[0007] S1, using an acceleration sensor to collect vibration signals of the bearing under different working conditions to obtain a plurality of 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 an improved inverted pyramid deep separable convolution sequence feature extraction model to extract the time-frequency image to obtain a multi-dimensional feature vector;

[0010] S4, utilize the feature vector as input data, based on the Swin Transformer deep learning network to train the model, to establish a bearing fault diagnosis model capable of adapting 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] Through the above technical solutions, compared with the prior art, the present application has the following technical advantages:

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

[0014] (2) For the noise interference problem commonly encountered in industrial applications, the method of the present application can effectively extract useful fault features from noise. Through the design of grouped convolution and dynamic separable blocks, the model improves the extraction ability of bearing fault feature information while maintaining lightweight, enhancing the robustness of the model in complex noise environments.

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

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0017] Figure 1 The bearing fault noise-resistant diagnosis method flowchart based on IPDSCS and Swin Transformer provided by the present application.

[0018] Figure 2 The deep separable convolution structure diagram provided by the present application.

[0019] Figure 3 The residual depth separable block structure diagram provided by the present application.

[0020] Figure 4 The dynamic separable block structure diagram provided by the present application.

[0021] Figure 5 The Swin-Transformer structure diagram provided by the application.

[0022] Figure 6 The bearing fault detection model structure diagram based on IPDSCS and Swin Transformer provided by the application.

[0023] Figure 7 The RGB three-channel learning schematic diagram for grouping convolution to fault TF image provided by the application.

[0024] Figure 8 The diagnostic flowchart of the IPDSCS-SwinT method provided by the application.

[0025] Figure 9 The test bench schematic diagram of the CWRU data set in embodiment one.

[0026] Figure 10 The class activation mapping thermograph of the proposed method IPDSCS-SwinT provided by the application. DETAILED DESCRIPTION

[0027] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0028] The embodiment of the application discloses a bearing fault noise resistance diagnosis method based on IPDSCS and Swin Transformer, referring to Figure 1 The method comprises the following steps S1-S5:

[0029] S1, collecting vibration signals of the bearing under different working conditions by using an acceleration sensor to obtain a plurality of groups of original vibration signal data;

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

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

[0032] Among them, wavelet transform is a powerful signal analysis tool that combines time and frequency analysis, especially suitable for processing non-stationary signals and capturing local features in signals. The core advantage of this analysis method is its multi-resolution capability, that is, it can provide detailed information of the signal at different scales. Wavelet transform changes the size and shape of the window instead of being fixed, so it can adaptively focus on different parts of the signal. With the adjustment of time window and frequency window, wavelet transform can achieve fine time resolution in high frequency area and good frequency resolution in low frequency area. Through the scaling and translation operation of wavelet transform, the signal is gradually refined at multiple scales, so as to focus on any details of the signal. The wavelet transform signal f(t) is the result of summing the product of the scaled and translated wavelet function (ψ(·)) over the entire period where the signal exists, as shown in equation (1).

[0033]

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

[0035] Among the many wavelet basis functions, cmor3-3 wavelet stands out because of its unique properties, showing less variability, which makes it widely used in the field of signal processing. As a single-frequency complex sinusoidal modulation Gaussian wave, cmor wavelet not only performs well in time domain, but also has excellent characteristics in frequency domain. This wavelet can provide high-resolution time-frequency analysis when processing signals, especially 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 outstanding ability of cmor3-3 wavelet to maintain high resolution of signal characteristics, which 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 more effectively extracted from noise, providing strong support for intelligent fault diagnosis of rolling bearings. Its expression is shown in equation (2).

[0037]

[0038] Where 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. In this embodiment, F c = F b = 3.

[0039] S3, a feature extraction model based on an improved inverted pyramid deep separable convolution sequence is constructed to extract the time-frequency image to obtain a multi-dimensional feature vector;

[0040] The improved inverted pyramid deep separable convolution sequence comprises:

[0041] A grouped convolution layer is configured to perform a grouped convolution operation on the time-frequency image.

[0042] A residual deep separable block is configured to perform feature extraction on the output of the grouped convolution layer.

[0043] A dynamic separable block is configured to perform feature extraction on the output of the residual deep separable block.

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

[0045] The size of the ordinary convolution is WH and C channels, and the input of the grouped convolution can be divided into n groups: Each convolution kernel is also divided into n groups: The output feature map size is W'xH'xg. Therefore, the parameter quantity and operation quantity of the grouped convolution are shown in formulas (3) and (4).

[0046]

[0047]

[0048] In the formula, params represents the parameter quantity of the grouped convolution; FLOPs represents the operation quantity of the grouped convolution; k represents the number of convolution kernels in each group; n represents the grouping number; C represents the input channel number; W', H' represents the width and height of the output feature map.

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

[0050] In addition, Global Depthwise Convolution (GDC) is introduced in each channel. GDC is a special convolution operation, where the number of groups G is equal to the number of input and output channels N and C, 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, enabling adaptive learning of different regions of the vibration signal.

[0051] 2). Residual Depthwise Separable Block

[0052] Depthwise Separable Convolution is an efficient convolution technique used in Convolutional Neural Networks (CNN), which reduces the number of parameters and computational burden of the model by simplifying the convolution operation, thereby improving the running efficiency and lightweight level of the model. As shown in Figure 2 Depthwise Convolution, which applies a dedicated convolution kernel (Filters) to each input channel (Channel Input) independently. This means that the convolution operation of each channel is performed separately, without information integration between different channels. This method makes the convolution operation only performed in the dimension of a single channel, thereby significantly reducing the number of parameters and computational load. Then, Pointwise Convolution uses a 1x1 convolution kernel, which integrates the features of different channels based on the output of Depthwise Convolution, achieving cross-channel information interaction. In this way, Pointwise Convolution fuses the information of each channel to form a more comprehensive image feature representation (Maps). Since Depthwise Convolution and Pointwise Convolution can be performed independently, Depthwise Separable Convolution significantly reduces the number of required parameters, which not only reduces the computational burden of the model, but also makes the model more lightweight compared to traditional convolution operations.

[0053] The present embodiment proposes a new residual depthwise separable module, as shown in Figure 3As shown, Residual Network (ResNet) is a deep neural network architecture. It can easily scale to very deep layers without performance degradation. This design makes ResNet highly effective for handling large-scale image datasets and performing complex tasks. The residual deep separable module integrates deep separable convolution techniques into the residual blocks, extracting features through 1x1 separable convolutions in skip connections. This not only enhances gradient backpropagation but also ensures differentiation between input and output, thereby improving the network's expressive power. Furthermore, to further improve module performance, Layer Normalization (LN) and the GeLU activation function are used to construct the residual blocks. LN normalizes each feature of each sample, making the model insensitive to variations in the distribution of input data, thus making it more stable when handling data with different distributions. Layer normalization is independent of batch size, making it particularly suitable for variable-length sequence data and effectively extracting specific features from faulty samples. By statistically analyzing the inputs of all neurons in each layer on each training sample and normalizing the inputs using the resulting mean and variance, the generalization ability and robustness of the model can be improved.

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

[0055]

[0056]

[0057] H represents the number of hidden neurons within the layer, where It is the sum of the normalized inputs of the i-th hidden unit in the l-th layer. 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 while preserving the input information. GeLU is shown in Equation (7).

[0059] GeLU(x)=x*P(X,,x)=x*Φ(x) (7)

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

[0061]

[0062] where μ and σ represent the mean and standard deviation of the Gaussian distribution, respectively. Since this function cannot be directly computed, the GeLU function can be approximated 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 throughout the entire real number range, and it will not saturate throughout the entire real number range, which means that its derivative will not tend to zero. This feature is crucial for neural network training using optimization algorithms such as gradient descent. Because the non-saturated activation function can guarantee the stable propagation of the gradient, thereby improving the efficiency and stability of network training. In contrast, ReLU, although it performs well on positive inputs, its derivative is zero on negative inputs. This can cause the gradient to disappear during backpropagation, affecting the learning efficiency of the network. In addition, ReLU is a piecewise linear function, which is not smooth in the processing of positive and negative inputs, which may introduce instability in the optimization process. The smoothness of GeLU helps to alleviate these problems, making the optimization process more stable and improving the network's ability to learn complex function mappings.

[0066] 3). Dynamic separable block

[0067] Dynamic convolution is a variant of convolutional neural networks (CNN). It introduces adjustable dynamic weights, allowing the convolution kernel to adaptively adjust parameters according to the characteristics of the input data. This dynamic adjustment mechanism gives the network greater adaptability compared to traditional fixed weight convolution kernels, allowing it to respond more flexibly to different input situations.

[0068] In order to build a model that is both lightweight and can effectively extract bearing fault features, this embodiment designs a novel dynamic separable block. This structure is shown in Figure 4 It relies on attention mechanisms and can dynamically aggregate multiple convolution kernels into a new separable convolution weight matrix according to input data. As shown in formula (11), this process essentially increases the number of channels, expanding them 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 perceiver used to learn the importance weight of the input feature; x represents an input feature map, representing an original signal or a feature after grouped convolution; W represents a set of convolution kernels, containing K different convolution kernels. b represents a bias term used to adjust the output of the convolution kernel; pi represents an attention weight used to control the degree of fusion of different convolution kernels; k is an index representing different convolution kernels.

[0072] where y is a static perceiver, K convolution kernels {W k} and attention weights {pi k (x)} linearly aggregate. The introduction of the dynamic separable module is 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, thus avoiding the increase in the number of network parameters. The purpose is to realize an adaptive receptive field adjustment mechanism. When processing two-dimensional wavelet transformed input data, this module can dynamically adjust its receptive field according to different location and scale features 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 extraction ability of the data features. In addition, since the dynamic separable block design focuses on working without expanding the channel depth, it helps to control the number of network parameters, thereby reducing the risk of overfitting. This design not only allows the convolution kernel to more flexibly adapt to the changing input characteristics, but also enhances the adaptability and generalization ability of the entire network.

[0073] S4, using the feature vector as input data, model training based on 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 attention mechanism. It relies on attention mechanism and can efficiently process large image data. As shown in Figure 5 , it includes:

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

[0076] Layer normalization layer for normalizing the output of the windowed multi-head self-attention mechanism;

[0077] MLP layer for feature extraction of 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, aiming to improve computational efficiency and storage efficiency. By using multiple parallel self-attention heads, SwinTransformer can learn data features from different angles simultaneously. Each self-attention head focuses on a subset of input data, i.e., a window, which allows 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 input data. The windowed multi-head self-attention mechanism is a flexible neural network design that allows the model to balance attention on global and local information according to the processing task. Smaller windows can capture more local details, while larger windows can include more extensive data information, 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 equation (12).

[0079]

[0080] wherein, is the query, key, and 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 represents the relative position between patches; d k represents the dimension of the query, key, and value.

[0081] In this embodiment, the extracted feature vector is input into the Swin Transformer network for model training. For example, the model contains 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 training, the test data is input into the model, and the accuracy rate 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, which converts the signal 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 designed to extract bearing fault features in noisy environments. Through 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 performed, and it is optimized accordingly.

[0084] This method combines the inverted pyramid deep separable convolution sequence (IPDSCS) with Swin Transformer to form an efficient fault diagnosis model. The core of the model is two innovative network structures: residual deep 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's parameter size and complexity controllable.

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

[0086] On the other hand, the dynamic separable block introduces an adaptive mechanism that adjusts the convolution filter according to the number of input experts, allowing the network to focus more on information features and less on noise. This dynamic adjustment provides the model with a certain degree of flexibility, which is crucial for handling changes 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 more finely process large-size image data and capture global and local features in image or sequence data, thereby improving the accuracy and robustness of fault diagnosis.

[0088] Reference Figure 6As shown, the method serially combines a pyramid depth separable convolution sequence with a Swin-Transformer model, and the parameter design is shown in Table 1. The core goal of the design of the method is to solve the limitations of traditional bearing fault diagnosis methods in feature extraction and model construction. 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 through its pyramid structure and depth separable convolution technology, realizing multi-scale feature extraction. The Swin Transformer model effectively processes sequence data and captures long-range dependencies in the data through its attention mechanism-based design. 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 fully utilize the strengths of both, achieving deeper feature learning and more accurate fault pattern recognition.

[0089] First, the pyramid depth separable convolution sequence is a new type of feature extraction module that combines pyramid structure and depth separable convolution technology. The pyramid structure can capture different details and features of the input data at different levels, enabling multi-scale feature extraction. As shown in Figure 7 As shown, the grouped convolution in different scales groups the wavelet-transformed image by RGB channels, with the RGB three groups of channels representing the features of the image in the red, green, and blue channels, respectively. Grouped convolution learns features independently on each channel, and by performing convolution operations on each group of channels separately, the network can better capture feature changes in different color channels of the image, thus more accurately understanding the image content. Convolution operations on each channel are performed independently, so different channel features can be processed in parallel, speeding up the model's training and inference speed. Grouped convolution helps better utilize image features, improve model expressiveness, reduce computational complexity and parameter quantity, and sends features of different channels to different depth separable convolution blocks in the subsequent stage. Depth separable convolution decomposes convolution operations into depth convolution and point-wise convolution, effectively reducing the model's parameter quantity and computational complexity. Residual connection allows the model to directly learn the residual between input and output, helping to alleviate the gradient vanishing problem and speed up the training process.

[0090] Then, the dynamic separable block can dynamically adjust the number of convolution kernels, so that the model has stronger flexibility and adaptability. The dynamic separable block introduces a dynamic convolution mechanism, and sets the dynamic channel number to 3 by default, 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 convolution kernel weight 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 model overfitting 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 bearing fault noise-resistant diagnosis. Secondly, Swin-Transformer is a neural network structure based on self-attention mechanism, which 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 image or sequence data, improving the modeling ability and generalization performance of the model.

[0091] IPDSCS-Swin Trans can fully utilize the advantages of both, achieve stronger feature extraction and model modeling, and further improve the accuracy and stability of bearing fault diagnosis, and realize end-to-end fault diagnosis.

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

[0093]

[0094]

[0095] In summary, the present application proposes a bearing fault diagnosis method combining inverted pyramid deep separable convolution sequence (IPDSCS) and SwinTransformer. The process framework of the IPDSCS-SwinT method is shown in Figure 8 , and the diagnostic procedure is summarized as follows.

[0096] 1. Through end-to-end one-dimensional data acquisition, complete information can be directly obtained 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 conversion, the original one-dimensional data can be analyzed in time and frequency, and converted into a TF image. This conversion analyzes the data in both time and frequency dimensions, allowing us to more comprehensively understand the characteristics and variation rules of the data.

[0098] 3. When performing wavelet transform, a noise 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 truly reflect the situation of the data.

[0099] 4. Fault diagnosis and anti-noise fault diagnosis are realized by analyzing and processing the TF image data. The TF image is classified and identified, so as to judge whether the data has abnormality or fault.

[0100] 5. The diagnosis result can be displayed through visualization and analysis. By presenting the diagnosis result in the form of image or chart, the state and fault condition of the data can be intuitively analyzed.

[0101] The following is an embodiment to verify the advantages of the IPDSCS-SwinT method proposed in the application on the benchmark bearing fault diagnosis data set. The benchmark data set is the Case Western Reserve University (CWRU) bearing data set. The IPDSCS-SwinT method of the application is written using the TensorFlow 2.5.0 framework in Python 3.8. All experiments are carried out under the same running environment, and the computer configuration includes Windows 10 16GB RAM, i9-9900K CPU and NVIDIA RTX 2070SUPER 8GB graphics processor.

[0102] Fault diagnosis of different methods

[0103] To verify the superiority of the method proposed in the application, the following deep learning methods are compared. In order to reasonably compare the characteristics of the methods, the parameters of all comparison models are consistent with the original text as much as possible. The detailed description of the comparison methods 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 is stacked by multiple convolutional layers and pooling layers, and introduces the Dropout regularization technology. By randomly setting the output of part of the neurons to zero, the risk of overfitting is reduced, and the local response normalization layer is used after the activation of the subsequent layer, which is used to suppress the activity of neurons and enhance the robustness of the model. Moreover, AlexNet first widely uses ReLU (Rectified Linear Unit) as the activation function, and GeLU is selected as the activation function for comparison in the embodiment.

[0105] b.InceptionV3: InceptionV3 is a deep convolutional neural network architecture proposed by Google Research team in 2015. InceptionV3 uses a structure called Inception module, which is an efficient feature extraction module. It processes the input in parallel with different scales and sizes of convolution kernels, and then concatenates the results in the depth direction. Auxiliary classifiers are introduced in the middle layers to help the gradient backpropagation during training. And depthwise separable convolution is introduced in the network, which is a lightweight convolution operation. Finally, global average pooling layers are used instead of fully connected layers to reduce the number of parameters and reduce the risk of overfitting. InceptionV3 and the inverted pyramid structure in the method of the invention are both a variant of the Inception structure, which is necessary for comparison.

[0106] c.VGG16: VGG16 is a deep convolutional neural network model proposed by Visual Geometry Group of Oxford University in 2014, which is composed of 16 layers of neural networks. VGG16 uses consecutive convolution layers to extract features, and has a deeper network structure than previous models. Each convolution layer uses a small size of 3x3 convolution kernel and convolution operation with stride 1, which helps to increase the non-linear representation ability of the network. VGG16 uses max-pooling layers to reduce the spatial dimension of the feature map and preserve important feature information. After the convolution layer, several fully connected layers are connected to map the high-level features to the class labels. The network structure of VGG16 is relatively simple and clear, the size of the convolution kernel and the stride are consistent, and the feature map size is halved after each pooling.

[0107] d.SwinTransformer: Swin Transformer is a new type of Transformer architecture proposed by Microsoft Asia Research Institute 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 processed using a windowed approach, which avoids the high memory and computational complexity of traditional global attention calculation. Local attention calculation is 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 employs an efficient network design that reduces the number of parameters and computational cost through depthwise separable convolutions and channel shuffle operations. ShuffleNet uses depthwise separable convolution operations, which decompose standard convolutions into two steps: depthwise convolution and pointwise convolution. It also introduces channel shuffle operations, which shuffle the features across different channels, increasing the communication and information transfer between features.

[0109] In the process of comparative experiments, we use 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 the learning rate is set to 0.0001, which is a relatively small value, which helps the model to converge stably in the training process. To ensure the reliability of the experimental results and reduce the influence of accidents, we implement 10-fold cross-validation technology. In addition, we monitor the change of loss value in the training process. When the loss value tends to be stable, that is, the performance of the model no longer improves significantly, we choose to stop iteration to avoid the occurrence of overfitting phenomenon.

[0110] Case: CWRU Bearing Dataset

[0111] We choose the CWRU bearing dataset as one of the benchmark datasets. This dataset is collected from a fault experiment on a SKF6205-2RS rolling bearing, and the test bench is like Figure 9The test rig consists of fan end and drive end rolling bearings, test base, torque transducer and electric motor. We select the drive end rolling bearing data as experimental data. At the sampling frequency of 12 kHz, we use different four motor load to drive the faulty bearing, and collect the vibration acceleration signal of the bearing through the acceleration sensor. We mainly use artificial electrical spark processing method to construct single-point groove faults with diameters of 0.007 inches, 0.014 inches, 0.021 inches and 0.028 inches respectively at three positions of the bearing ball (BA), inner ring (IR) and outer ring (OR). We select the fault data under the conditions of 1 hp, 1772 r / min and 2 hp, 1750 r / min in the experiment to set three data sets, each of which contains 4 kinds of ball faults, 4 kinds of inner ring faults and 3 kinds of outer ring faults. The sample classification of the data set is shown in Table 2. We use random sliding window sampling, and the sliding window is 1024, that is, each sample has 1024 data points. Sliding window sampling can preserve the time sequence information of time series data. Because it slides on the data according to a fixed step, it can obtain rich time domain features, frequency domain features and time-frequency features, and each sampling considers the relationship between the previous and next data points without additional data processing cost. The ratio of training samples to test samples in each data set is 7:3.

[0112] Table 2 CWRU rolling bearing data set sample classification

[0113] Class Failure type Failure size 0 Rolling body 0.007 1 Rolling body 0.014 2 Rolling body 0.021 3 Rolling body 0.028 4 Inner ring 0.007 5 Inner ring 0.014 6 Inner ring 0.021 7 Inner ring 0.028 8 Outer ring 0.007 9 Outer ring 0.014 10 Outer ring 0.021

[0114] 1 Performance under strong noise

[0115] In industrial production, the vibration signal generated by the bearing during operation is often disturbed and affected by environmental noise. This noise may mask the fault information, making 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 (SNRs). The definition of signal-to-noise ratio is as follows:

[0116]

[0117] In the formula, P signal is the original signal power, and P noise is the noise power. According to the formula, SNR is inversely proportional to P noise , so the more serious the signal is contaminated by noise, the smaller the value of SNR.

[0118] We select the data set under the operating condition of 1750 r / min and 2 hp load for the experiment, and add 6, 3, 0, -3 and -6 dB simulated industrial noise to the data set. First, if the number of noise samples is sufficient, we can increase the signal-to-noise ratio in both training and testing samples. The purpose of this is to enable the model to better adapt to the high-noise environment during training, while accurately identifying fault signals and resisting noise interference during the testing phase. Second, for samples collected under noisy conditions, we only add signal-to-noise ratio to the testing samples. This is to simulate the actual working environment, ensuring that the model can effectively operate and diagnose faults in a real noisy environment. Finally, for limited samples collected under normal conditions, we only add signal-to-noise ratio to the training samples. This method helps the model learn fault features under the influence of noise during training, improving the model's robustness and generalization ability.

[0119] We add the signal-to-noise ratio in both training and test samples to simulate the ubiquitous noise in real working environment and evaluate the performance of the model in the presence of ubiquitous noise. Table 3 is the fault diagnosis rate of each comparative method under 5 signal-to-noise ratios. Under the condition of 6dB signal-to-noise ratio and 300 samples, the single bearing fault of each comparative method can reach 100%. However, as the number of samples decreases, it can be seen that the fault diagnosis effect of IPDSCS-SwinT is remarkable. VGG16 still has a good fault diagnosis rate and can have certain noise resistance in the TF image of bearing fault. Under the conditions of 3dB and 0dB: IPDSCS-SwinT maintains the highest accuracy, 99.93% and 99.30% respectively, showing the best performance under moderate noise level. VGG16 and InceptionV3 also have good performance, with accuracy slightly lower than IPDSCS-SwinT, but still maintaining around 99%. Even in the case of lower signal-to-noise ratio and greater noise influence, VGG16 can still maintain a very high accuracy. As the signal-to-noise ratio further decreases to -3dB and -6dB, IPDSCS-SwinT still performs best, with accuracy of 98.89% and 98.63% respectively. Swin Transformer and VGG16 also perform relatively well under the condition of -3dB, but as the signal-to-noise ratio further decreases, the accuracy decreases significantly. Under the condition of -6dB, 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 signal-to-noise ratio conditions, but its accuracy decreases rapidly under low signal-to-noise ratio conditions, indicating that it is more sensitive to noise. VGG16 and InceptionV3 also perform well under moderate signal-to-noise ratio conditions, but their performance is not as good as IPDSCS-SwinT in high noise environment. Swin Transformer has high accuracy under moderate signal-to-noise ratio conditions, but its performance decreases rapidly under extremely low signal-to-noise ratio. AlexNet and ShuffleNet have high accuracy under high signal-to-noise ratio conditions, but their accuracy decreases significantly when the noise is large. IPDSCS-SwinT combines the advantages of inverted pyramid depth separable convolution sequence and Swin Transformer, and can effectively extract key features from vibration signals and achieve accurate fault diagnosis in the presence of noise. IPDSCS-SwinT exhibits excellent performance under various signal-to-noise ratio conditions, especially in terms of stability and accuracy under low signal-to-noise ratio conditions, highlighting the outstanding robustness and generalization ability of the method under different noise environments.

[0120] Table 3 Single noise fault diagnosis rate

[0121]

[0122] Table 4 shows the fault diagnosis accuracy of different deep learning methods on 200 samples without noise in different signal-to-noise ratios. At 6dB, the accuracy of AlexNet is 99.16%, and as the signal-to-noise ratio decreases, the accuracy gradually decreases, and at -6dB, it decreases to 86.67%. This shows that the resistance of AlexNet to noise decreases as the signal-to-noise ratio decreases. 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 decreases to -6dB. This shows that VGG16 has good stability at different noise levels. InceptionV3 performs similarly to VGG16 at high signal-to-noise ratios, but its accuracy decreases faster at low signal-to-noise ratios. 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 signal-to-noise ratio decreases, the accuracy decreases to 87.23% at -6dB. Swin Transformer has an accuracy of 98.56% at 6dB, and decreases to 83.54% at -6dB. Swin Transformer performs well at high signal-to-noise ratios, but its performance decreases significantly at low signal-to-noise ratios. IPDSCS has an accuracy of 98.08% at 6dB signal-to-noise ratio, and 84.24% at -6dB. The performance of IPDSCS remains relatively stable in the medium to low signal-to-noise ratio range. IPDSCS-SwinT shows the best fault diagnosis accuracy in all tested signal-to-noise ratios, from 99.62% at 6dB to 90.73% at -6dB. Even at low signal-to-noise ratios, the accuracy of IPDSCS-SwinT decreases less, showing the best robustness.

[0123] Table 4 Fault diagnosis rate of test samples without 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 sample quantities without noise in the training samples. AlexNet has a diagnosis accuracy of 96.71% at 6dB and 83.70% at -6dB. This indicates that the greater the noise interference, the more severe the decline in diagnosis performance of AlexNet. VGG16 shows good robustness, with a smaller decline in accuracy 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 they are less resistant to noise when the sample quantity is small. ShuffleNet and Swin Transformer have high accuracy under high signal-to-noise ratio conditions, but their performance decreases significantly under very low signal-to-noise ratio conditions, 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 exhibits the highest fault diagnosis accuracy under all signal-to-noise ratio conditions, from 98.07% at 6dB to 86.29% at -6dB. Even under conditions of small sample quantities and low signal-to-noise ratios, the accuracy of IPDSCS-SwinT decreases less. IPDSCS-SwinT still exhibits excellent performance under various signal-to-noise ratio conditions even with a small sample quantity, and its ability to maintain accuracy under low signal-to-noise ratio conditions indicates that the model has very strong robustness.

[0127] Table 5 Fault diagnosis rate without noise in training samples (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) heat maps for visualization to show the areas that IPDSCS-SwinT focuses on when making a diagnosis, such as Figure 10The CAM can clearly highlight the signal features related to the fault and maintain this ability in a noisy environment, which will demonstrate its significant advantage in the bearing fault diagnosis task. The CAM can accurately indicate the specific frequency band or time domain feature where the fault occurs, which indicates that the IPDSCS-SwinT method has the ability to accurately locate the fault type. In the presence of noise and other interference, the CAM can still highlight the relevant fault features, which will prove that the IPDSCS-SwinT model has good robustness to noise and interference in practical applications. The heatmap of IPDSCS-SwinT can consistently highlight the signal features corresponding to the known fault patterns, which indicates that the model can learn the complex relationships related to bearing faults.

[0130] 2. Performance under complex conditions

[0131] In the actual industrial environment, the complex situation of bearing operation is often affected by multiple factors, including sudden changes in speed and load, and the presence of noise. In order to more comprehensively study the influence of these factors on the performance of the bearing, 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. First, sudden changes in speed can cause changes in bearing force and vibration frequency, affecting the working state and performance of the bearing. Second, sudden changes in load are also common in industrial environments, which can cause changes in the load distribution of the rotor and the force borne by the bearing, thereby affecting the operation of the bearing. Finally, these complex situations are often accompanied by the presence of noise, which can come from the operation of the bearing itself, external environmental interference, and other factors. Therefore, we considered the influence of noise on the performance of the bearing in the experiment and took appropriate measures to reduce the influence of noise to more accurately evaluate the working state and performance of the bearing. Through these experiments, we can better understand the working characteristics of the bearing under different working conditions and provide reliable bearing fault diagnosis and prediction for industrial production.

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

[0133] Precision: also known as precision, is an evaluation indicator for predicted results. It refers to the proportion of actual positive samples in the predicted positive samples.

[0134]

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

[0136] Precision: It refers to the proportion of samples that are correctly predicted as positive classes among all samples that are actually positive classes.

[0137]

[0138] False Positive Rate: Also known as the false detection rate, it refers to the ratio of the number of samples classified incorrectly to the total number of samples, and 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 evaluating models on imbalanced datasets.

[0141]

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

[0143] Table 6 reflects the evaluation of the diagnostic performance of each method under different signal-to-noise ratio (SNR) conditions using evaluation indicators when the sample size is relatively small. AlexNet and VGG16 perform similarly on most evaluation indicators, but are slightly lower than IPDSCS-SwinT in F1 score. This indicates that when considering precision and recall comprehensively, their performance is slightly inferior. InceptionV3 performs the worst among all models, especially in false positive rate, which may indicate that this model is more prone to false positive predictions in fault diagnosis. Swin Transformer and ShuffleNet also exhibit 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 exhibits the best or near-best performance on all evaluation indicators. In particular, in precision and F1 score, this indicates that IPDSCS-SwinT has the highest accuracy and comprehensive performance in the fault diagnosis task when the sample size is 80.

[0144] Table 6 1772-1750 variable load noise (variable operating condition) fault diagnosis evaluation index (80 samples)

[0145]

[0146]

[0147] In the analysis of the CWRU bearing dataset under the 1772-1750 variable load noise condition, we used the t-SNE technique to visualize the fault feature representation of 200 samples at a signal-to-noise ratio of 0 dB. These fault types were classified by 7 different deep learning methods. Among these methods, AlexNet, VGG16 and ShuffleNet showed interference in their feature representations when distinguishing between the outer ring 21-inch fault and the inner ring 28-inch fault, which may be due to the fact that the signal characteristics of these two fault types are too close in high-dimensional space, making it difficult for the model to distinguish. Similarly, all 7 methods also showed mutual interference in the feature information of the spherical 14-inch fault and the outer ring 14-inch fault, further exacerbating 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 remained well-structured even at a low signal-to-noise ratio of 0 dB, which means that the model can more effectively distinguish different fault types even in noisy conditions. The excellent performance of the IPDSCS-SwinT model in t-SNE visualization proves 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 compound noise

[0149] In industrial environments, the operating state of bearings is often disturbed by multiple noises. These noises can 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 situation in industrial environments, we selected the dataset with a speed of 1772r / min and a load of 1hp for experiments. By adding different intensities of Gaussian white noise, we can simulate the influence of different sources of noise on the bearing signal in industrial environments, thus more comprehensively evaluating the diagnosis state and performance of the model for bearing faults.

[0150] In the diagnosis of bearing faults under complex noise, we evaluated the performance of various comparative methods through the confusion matrix. The results show that, under variable noise and variable load conditions, the VGG16 model performs well. However, in the case of complex conditions with multiple feature couplings and complex noise, its diagnostic performance declines. At the same time, InceptionV3, VGG16, ShuffleNet, and IPDSCS models generally perform poorly in terms of anti-interference performance under complex noise conditions. In contrast, AlexNet, although it performed only moderately in Section 2 above, exhibits stronger fault feature decoupling capabilities than VGG16 under complex noise environments. However, when compared with AlexNet, Swin Transformer, and IPDSCS-SwinT, IPDSCS-SwinT has a more significant advantage by introducing an attention mechanism that more effectively learns multiple feature representations under complex noise. The experiments verify that the attention mechanism of Swin Transformer provides a significant performance improvement in handling multiple feature coupling problems. Therefore, the IPDSCS-SwinT model exhibits excellent capabilities in multiple feature decoupling under complex noise. It not only accurately distinguishes fault signals that interfere with each other, but also exhibits low error rates in the confusion matrix, indicating that the model can effectively suppress noise and extract useful fault features. Even in noisy environments, IPDSCS-SwinT maintains good diagnostic capabilities. In addition, the IPDSCS-SwinT model maintains good performance in the confusion matrix under different noise conditions.

[0151] The bearing fault diagnosis method based on the inverted pyramid deep separable convolution sequence (IPDSCS) and Swin Transformer proposed by the present application has achieved remarkable results in terms of accuracy and robustness in bearing fault diagnosis. Experiments on the CWRU bearing dataset show that even in low signal-to-noise ratio and complex noise environments, the IPDSCS-SwinT method can maintain a high diagnostic accuracy rate. In addition, compared with other deep learning methods, IPDSCS-SwinT performs well in fault decoupling under multiple feature coupling and complex noise conditions. Furthermore, IPDSCS-SwinT can maintain stable diagnostic performance even with a small number of samples, demonstrating its robustness to changes in sample size.

[0152] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Therefore, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for diagnosing bearing fault noise immunity based on IPDSCS and Swing Transformer, characterized in that, Includes the following steps: S1. Use an accelerometer to collect vibration signals of the bearing under different operating conditions to obtain multiple sets of raw vibration signal data; S2. Perform wavelet transform on the original vibration signal data to convert it into a time-frequency image; S3. Construct an improved inverted pyramid depth-separable convolutional sequence (IPDSCS) feature extraction model to extract multi-dimensional feature vectors from the time-frequency image; wherein, the improved inverted pyramid depth-separable convolutional sequence (IPDSCS) includes: Grouped convolutional layers are used to perform grouped convolution operations on time-frequency images; Residual depth separable blocks are used for feature extraction from the output of grouped convolutional layers; Dynamically separable blocks are used to extract features from the output of residual depth separable blocks. S4. Using the feature vector as input data, the model is trained based on the Swing 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 sets of raw signal data include: Vibration signals corresponding to normal condition, 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 the cmor3-3 wavelet, as shown in the following formula: Among them, F c F is the center frequency of the wavelet. b is the bandwidth parameter; i is a fictitious unit; x represents the original vibration signal data.

4. The method according to claim 1, characterized in that, The grouped convolutional layer performs grouped convolution operations on the RGB channels of the time-frequency image, including the following: The input to the grouped convolution is divided into n groups, each containing a different feature map. Each convolutional kernel is also divided into n groups: The output feature map has dimensions W′×H′×g; W and H represent the width and height of the input feature map. The number of parameters and computational complexity of grouped convolution are shown in equations (3) and (4): In the formula, params represents the number of parameters in the grouped convolution; FLOPs represents the computational cost of the grouped convolution; k represents the number of kernels in each group; n represents the number of groups; C represents the number of input channels; and W' and H' represent the width and height of the output feature map.

5. The method according to claim 1, characterized in that, The residual depth separable block includes the following: Depthwise separable convolutional layers are used to perform depthwise separable convolution operations on input features; Pointwise convolutional layers are used to perform pointwise convolution operations on the output of depthwise separable convolutional layers. Residual connections are used to residually connect the outputs of depthwise separable convolutional layers with the outputs of pointwise convolutional layers.

6. The method according to claim 5, characterized in that, The residual connections employ layer normalization and the GeLU activation function. Layer normalization: By statistically analyzing the inputs of all neurons in each layer on each training sample, the mean and variance obtained are used to normalize the inputs. The mean is shown in formula (5), and the variance is shown in formula (6). In the formula, H represents the number of hidden neurons within the layer, where It is the sum of the normalized inputs of the i-th hidden unit in the l-th layer; Under layer normalization, all hidden neurons within the same layer have the same normalization terms μ and σ; The activation function GeLU is approximately represented by formula (9): x represents the input to the activation function GeLU.

7. The method according to claim 1, characterized in that, The dynamically separable block includes the following: Dynamic convolutional layers are used to dynamically adjust the weights of the convolutional kernels based on the input features; An attention mechanism layer is used to weight the output of the dynamic convolutional layer; it expands the output by a factor of K and applies dynamic attention to the K channel groups. y=g(W T (x)x+b(x)) (10) In the formula, y represents the static perceptron, which is used to learn the importance weights of the input features; x represents the input feature map, which represents the original signal or the features after grouped convolution; W represents the set of convolution kernels, which contains K different convolution kernels; b represents the bias term, which is used to adjust the output of the convolution kernels; π represents the attention weight, which is used to control the degree of fusion of different convolution kernels. k is an index representing different convolution kernels.

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

9. The method according to claim 8, characterized in that, The window-based multi-head self-attention mechanism layer adapts to different data characteristics and task requirements by adjusting the window size; the attention mechanism is shown in formula (12): in, The information to be queried, related information to the information to be queried, and specific information related to the information to be queried; M 2 The total number of patches in a window, d k 'b' represents the dimension of the query, key, and value; 'b' represents the relative positional deviation, used to characterize the relative position between patches.

Images