Flow field vortex structure identification method based on multi-scale feature fusion

Abstract

The invention discloses a flow field vortex structure recognition method based on multi-scale feature fusion. The method mainly comprises the steps of flow field data input and preprocessing, construction of a multi-scale feature fusion flow field vortex structure recognition network MFV-Net, design of a direction enhancement residual module OER-Block, design of a multi-channel feature enhancement module MCFE and design of a multi-scale feature fusion reconstruction module MSR. Simulated flow field data is used as input, a vortex structure in a flow field is identified through an MFV-Net network, and vortex points and non-vortex points in the flow field data are divided. The problems that in the prior art, the flow field vortex structure recognition precision is low, and a small-scale vortex structure is lost are solved. Three multi-scale vortex characteristic information enhancement modules are designed, and the three modules are integrated to construct a multi-scale characteristic fusion flow field vortex structure identification network MFV-Net, so that flow field global information is fully combined, the segmentation capability of the network on the vortex structure is improved, and accurate identification of the vortex structure in the flow field is completed.

Description

Technical Field

[0001] The present invention relates to the technical field of flow field feature recognition, and more specifically, to a flow field vortex structure recognition method based on multi-scale feature fusion technology. Background Art

[0002] Flow field structure identification is of great value in the fields of fluid mechanics, meteorology, oceanography, etc., and can reveal the basic characteristics and evolution laws of flow phenomena. The flow field structures mainly include vortices, shock waves, expansion waves and boundary layers. Among them, vortex structure is one of the most important characteristics of flow field characteristics, which implies more flow field information. The accurate identification of vortex structure is of great significance for studying flow field laws, understanding the potential physical mechanism of flow field and aircraft design. It has been highly valued by researchers and has been studied in depth in recent years. However, since there is no very clear mathematical definition of vortex structure, there are great limitations in the process of identifying vortex structure in flow field.

[0003] The identification methods of vortex structures in flow fields can be roughly divided into three categories: local detection methods, global detection methods and machine learning-based methods. Local methods are usually based on the physical properties of the flow field and are conceptually easy to explain, such as "local high vorticity" and "local low pressure". They only use local flow field information to calculate some criteria, so they can quickly obtain results, but the detection accuracy is low and the threshold is highly dependent. Global methods are usually based on the global topological properties of the flow field, such as winding angle and instantaneous vorticity deviation (IVD). These methods have extremely high accuracy, but the calculation is large and takes more time than local methods. Therefore, it is difficult to apply these methods to identify vortex structures in large-scale data sets. The machine learning-based method is currently a new type of flow field vortex structure identification method. By inputting flow field data into a neural network for training, the vortex structure in the flow field can be automatically learned and extracted. Among them, convolutional neural network (CNN) is a type of machine learning algorithm that has made breakthrough innovations in the field of object recognition in recent years. Due to the locality and translation invariance of convolution operations, CNN can effectively resist recognition difficulties such as translation, deformation and rotation. Therefore, it is generally believed that CNN is very suitable for vortex structure identification tasks.

[0004] In the current research, the existing flow field vortex structure recognition networks based on convolutional neural networks, such as Vortex-Net, mostly use simple multi-layer convolution pooling superposition encoding and decoding operations, which inevitably causes the loss of global context information of the flow field. The high-level semantic information of the flow field is not combined with the underlying geometric information, resulting in inaccurate vortex structure recognition results. For example, the performance degradation problem caused by the deepening of the network and the problem of easy loss of small-scale vortex structure information caused by continuous downsampling. Therefore, based on the existing research, a vortex structure recognition network based on multi-scale feature fusion (Multi-scale feature Fusion flow Field Vortex recognition convolution NeuralNetwork, MFV-Net) is proposed, and three multi-scale information enhancement modules are designed to fully combine the global context information of the flow field and complete the accurate recognition of the flow field vortex structure. Summary of the invention

[0005] The present invention provides a vortex structure recognition method based on multi-scale feature fusion technology, which mainly uses multi-scale feature fusion technology and combines flow field global context information to achieve accurate recognition of vortex structure.

[0006] The purpose of the present invention is to propose a flow field vortex structure recognition method based on multi-scale feature fusion, in view of the fact that the flow field vortex structure recognition network in the existing vortex recognition technology mostly relies on simple convolution layers and pooling layers to implement downsampling and upsampling operations on flow field data features, which inevitably causes a large amount of loss of flow field global context information and a series of precision loss problems. The technical solution is as follows:

[0007] (1) In order to retain more physical grid information of the flow field, the vorticity field that integrates the original velocity field and the physical grid information is used as input to further improve accuracy and generalization. Vorticity is a pseudovector. It is expressed as ω → And defined as the velocity field v → The rotation of , which describes continuous motion. The vorticity calculation formula in the Cartesian coordinate system is as follows:

[0008]

[0009] Among them, x, y, and z are the real coordinates of different axes in physical space. It can be calculated by the following formula:

[0010]

[0011] In the above equation, It can be expressed as (vi+1,j,k-vi-1,j,k) / 2Δξ, as shown in the following equation, where J represents the matrix Jacobian.

[0012]

[0013] In order to introduce global information into the vortex identification method, the training data is labeled using the vortex identification method based on IVD (instantaneous vorticity deviation). The calculation of IVD is based on the vorticity calculated in (1), specifically the absolute value of the difference between the vorticity at a certain point and the global average vorticity of the flow field. The calculation formula is as follows:

[0014] IVD(x,t)=|ω(x,t)-ω avg (x,t)|

[0015] Where ω represents the vorticity, ω avg represents the average vorticity, and the calculated IVD is used as the criterion for distinguishing vorticity.

[0016] (2) In a non-uniform grid, there is a physical distance value between two grid points, which is not convenient for direct subsequent processing. Therefore, a grid transformation strategy is used to convert the non-uniform grid in the three-dimensional physical space into a uniform grid in the three-dimensional computational space. Each grid point has relevant position information and other relevant field information in the Cartesian coordinate system. Then the grid boundary is processed. During training, the input is a local flow field block for each point, but the points near the boundary do not have enough adjacent blocks to obtain a local patch, so it is necessary to fill the grid with the properties of the original flow field to obtain the preprocessed data.

[0017] (3) Divide the training data set and the test data set. First, use 80% of the data as the training data set and the remaining 20% ​​as the test data set. 80% of the training data set is used as the training set, and the remaining 20% ​​is used as the validation set. Use the IVD labels calculated in step (1) as the training data.

[0018] (4) Normalize the data. First, use the Z-score equation to standardize the entire vorticity field. The equation is as follows:

[0019]

[0020] In the equation, x is the vorticity value of the flow field, and μ and δ are the mean and standard deviation of all vorticity values. In this way, a standardized vorticity with a mean of zero and a standard deviation of one is obtained. The Z-score equation eliminates the difference in the vorticity field, and then the sigmoid function is used to normalize the vorticity field between 0 and 1. In this way, the normalized vorticity field is obtained.

[0021] (5) Constructing a network model. The present invention proposes a convolutional neural network (MFV-Net) for flow field vortex structure recognition based on multi-scale feature fusion. Based on the U-Net network model, three multi-scale feature enhancement modules are designed to help the network fully combine the global context information of the flow field and improve the accuracy of vortex structure recognition. First, in order to solve the gradient vanishing problem, the directional enhancement residual module (OER-Block) is used to replace the convolution block of U-Net; secondly, in order to reduce the information loss in the encoding and decoding process, a multi-channel feature enhancement module (MCFE) is constructed at the jump connection of U-Net; finally, in order to capture multi-scale feature information and improve the network's recognition ability of vortex structure, a multi-scale feature fusion and reconstruction module (MSR) is introduced at the bottom of U-Net.

[0022] (6) Design loss function. In the computational flow field, the grid information is discarded, so the importance of all points is the same. However, in the physical flow field, points close to the boundary of the object shape are more important than other points, and the importance of the point is inversely proportional to the distance. In order to retain the characteristics of the physical flow field, the weighted cross entropy loss function is used. The loss function is as follows:

[0023]

[0024] Where N p represents the number of grid points in a flow field block, y i and are the true label and predicted label of the i-th point respectively. In order to introduce distance information into the network model, it is necessary to pre-calculate the weight map of each flow field. The weight of each point is calculated by the following equation:

[0025]

[0026] where x i is the i-th grid point in the flow field block, d(x i ,B) represents x i The minimum distance to the flow field boundary point.

[0027] (7) Sample flow field blocks from the training data set and train the network model. The input is the coordinate information of the flow field, the three components of velocity (U, V, W), pressure, vorticity, instantaneous vorticity deviation (IVD), and weight ω i Flow field patches of size 32×32×32 are randomly sampled from the training dataset as the input of MFV-Net.

[0028] (8) Post-processing: combining the position coordinate information of each flow field block output by the network to finally obtain the prediction result of the entire flow field. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1This is the MFV-Net network framework diagram.

[0030] Figure 2 Enhanced residual module graph for OER-Block direction.

[0031] Figure 3 This is the MCFE multi-channel feature enhancement module diagram.

[0032] Figure 4 Diagram of the MSR multi-scale feature fusion and reconstruction module. DETAILED DESCRIPTION

[0033] To make the technical solution of the present invention clearer, the multi-scale feature enhancement module of the present invention is further described in detail below with reference to the accompanying drawings:

[0034] The network framework diagram of the present invention is as follows Figure 1 As shown, the specific process is as follows:

[0035] First, the above steps (1) to (4) are the data preprocessing part. The following specifically introduces step (5) to build the network model:

[0036] (5) Constructing a network model. The present invention uses U-Net as the basic model and designs a 3D U-Net for vortex structure recognition, called MFV-Net. Three multi-scale feature enhancement modules are designed to help the network fully integrate the global context information of the flow field and improve the accuracy of vortex structure recognition. First, to solve the gradient vanishing problem, the directional enhancement residual module (OER-Block) is used to replace the convolution block of U-Net. Second, to reduce information loss in the encoding and decoding process, a multi-channel feature enhancement module (MCFE) is constructed at the jump connection of U-Net. Finally, to capture multi-scale feature information and improve the network's recognition ability of small-scale vortex structures, a multi-scale feature fusion and reconstruction module (MSR) is introduced at the bottom of U-Net.

[0037] The MFV-Net network framework is shown in the figure Figure 1 As shown, it consists of a contraction and expansion part, with two OER-Block convolution operations in each step. Downsampling is performed using a maximum pooling layer with a stride of 2×2×2 at the last layer of each step. In addition, the number of feature channels is doubled after each downsampling step. The expansion part includes a 2×2×2 deconvolution (DeConv) that reduces the number of channels by half, a concatenation with the corresponding feature map from the encoder after MCFE feature enhancement, and the same OER-Block convolution operation as the contraction part. To overcome overfitting, dropout is added after the last few convolution layers. In the last layer of the network, a 1×1×1 convolution is used to reduce the output channels.

[0038] (5a) First, the improved directional enhanced residual module (OER-Block) is used to replace the original U-Net convolutional block. Based on the traditional residual module, OER-Block is deeply customized and optimized for the flow field vortex identification task. This module adopts a multi-scale feature fusion strategy, and by integrating convolution kernels of different sizes such as 1×1, 3×3, and 5×5, it can capture multiple scale features from local details to global structures. This design can greatly improve the model's adaptability to the diversity of vortex shapes and sizes;

[0039] OER-Block introduces Direction-Aware Convolution, which not only considers the relationship between adjacent pixels in space, but also the directional relationship between pixels. Through directional filters, local directional features in the flow field, such as the direction of the velocity vector, can be captured, allowing OER-Block to better understand the directional characteristics of the flow field and enhance its sensitivity to vortex structures. This convolution operation helps capture key features in the flow field and further improve the accuracy of vortex identification.

[0040] The implementation of direction-aware convolution can be achieved by defining a directional convolution kernel G(θ) to extract features in a specific direction. The direction-aware convolution Y of the input feature map X can be expressed as:

[0041] Y=G(0)*X

[0042] Structural tensor calculation,The structural tensor J can be used to capture the directional information in the flow field and is defined as:

[0043]

[0044] Where Ix and Iy are the gradients of the input image in the x and y directions, which can be calculated using the following convolution kernel:

[0045]

[0046] OER-Block retains some conventional advantages of the residual module, such as identity shortcut connections to promote effective propagation of gradients, as well as Batch Norm layers and PReLU activation functions between convolutional layers to improve the convergence speed and generalization ability of the model.

[0047] (5b) By using OER-Block to replace the traditional convolution block, the gradient vanishing problem can be alleviated to a certain extent and the ability to extract vortex features can be improved. However, in the encoding and decoding process, if the encoding result is directly connected to the decoding result, the vortex feature information at the connection will still be lost. Therefore, a multi-channel feature enhancement module (MCFE) is constructed at the jump connection of U-Net. This module enhances the feature expression ability of the flow field vortex structure through deformable convolution and inter-layer feature interaction mechanism, and realizes efficient transmission of the encoding results by effectively integrating vortex characteristic information at different levels;

[0048] The structure of MCFE is shown in the attached Figure 3 As shown, its mathematical definition is as follows:

[0049] P i =MCFE(X i ,concat(DeformConv(X i-1 ),upsample(Y i+1 )))

[0050] Among them, DeformConv represents the deformable convolution operation, concat represents the concatenation of feature maps of different channels, upsample represents upsampling using transposed convolution, Xi, Yi represent the codec outputs at different stages, and Pi represents the operation result after MCFE.

[0051] To keep the dimensions of the output data of different channels consistent, MCFE first performs upsampling or downsampling operations. Then, the encoded outputs from different levels are spliced ​​to make full use of the feature information at different levels. Next, by combining the spatial and channel attention mechanisms, a channel attention feature map is generated through statistical analysis to further improve the expression of vortex features. Finally, the generated attention features are fused with the output feature map through residual connections. MCFE can reduce information loss in the encoding and decoding process and enhance the model's ability to recognize vortex features in the flow field.

[0052] (5c) Finally, in order to address the problem that small-scale vortex structures are easily lost after multiple convolution operations during flow field vortex identification and improve the segmentation performance of the network, the study introduced a multi-scale feature fusion and reconstruction module (Multi-Scale Feature Fusion and Reconstruction Module, MSR), which aims to improve the model's recognition effect on small-scale vortices. The main components of this module include the following steps:

[0053] First, multi-level feature extraction: Through a hierarchical network structure, vortex features are extracted from different scales to ensure that small-scale details are retained during the convolution process. Specifically, 1×1 convolution and depth-wise separable convolution (using convolution kernels of sizes 3, 5, and 7) are used for multi-level feature extraction; Feature weighting: After feature extraction, a learnable weight ωi is assigned to the feature map of each scale. These weights are dynamically adjusted through the training process to reflect the importance of the feature map in the vortex identification task. The calculation of weights can be based on feature statistics or through a small neural network scoring mechanism. Weighted fusion: The weighted feature maps are combined by weighted summation to form the final fused feature map Ffinal;

[0054]

[0055] Among them, Fi is the i-th feature map, and ωi is the corresponding weight.

[0056] In the fused feature map, a recalibration mechanism is introduced to extract global information through global average pooling and generate weights between channels to dynamically adjust the expression of features:

[0057] Fadjusted=Ffnal·r(W.AvgPool(Fhnal))

[0058] Among them, σ is the activation function and W is the weight obtained by the fully connected layer.

[0059] Finally, the deconvolution layer is used to restore the small-scale vortex details in the fusion feature map, further improving the recognition accuracy and the ability to identify small-scale vortices in the flow field. This step ensures that the information of the small-scale vortex structure is effectively preserved in multiple convolution operations.

[0060] (6) Design loss function. In the computational flow field, the grid information is discarded, so the importance of all points is the same. However, in the physical flow field, points close to the boundary of the object shape are more important than other points, and the importance of the point is inversely proportional to the distance. In order to retain the characteristics of the physical flow field, the weighted cross entropy loss function is used. The loss function is as follows:

[0061]

[0062] Where N p represents the number of grid points in a flow field block, y i and are the true label and predicted label of the i-th point respectively. In order to introduce distance information into the network model, it is necessary to pre-calculate the weight map of each flow field. The weight of each point is calculated by the following equation:

[0063]

[0064] where x i is the i-th grid point in the flow field block, d(x i ,B) represents x i The minimum distance to the flow field boundary point.

[0065] (7) Sample flow field blocks from the training data set and train the network model. The input is the coordinate information of the flow field, the three components of velocity (U, V, W), pressure, vorticity, instantaneous vorticity deviation (IVD), and weight ω i Flow field blocks of size 32×32×32 are randomly sampled from the training dataset as the input of MFV-Net. The input size of the network is independent of the size of the flow field.

[0066] (8) Post-processing: combining the position coordinate information of each flow field block output by the network to finally obtain the vortex structure prediction result of the entire flow field.

Claims

1. A flow field vortex structure recognition method based on multi-scale feature fusion, which is characterized by: Based on the U-Net network framework, three multi-scale feature enhancement modules are designed to fully combine the global context information of the flow field and accurately identify the vortex structure of the flow field, including the following steps: (1) In order to retain more physical grid information of the flow field, the original velocity field is combined with the physical grid information to generate the vorticity field as the network input. The vorticity field that integrates the original velocity field and the physical grid information is used as the input to further improve the accuracy and generalization. In order to introduce global information into the vortex identification method, the training data is marked using the vortex identification method based on IVD (instantaneous vorticity deviation). The calculation of IVD is based on the vorticity calculated in (1), which is the absolute value of the difference between the vorticity at a certain point and the global average vorticity of the flow field. The calculation formula is as follows: IVD(x,t)=|ω(x,t)-ω avg (x,t)| in, ω represents the vorticity, ω avg represents the average vorticity, using the calculated IVD as the criterion for discriminating vorticity; (2)(2) In a non-uniform grid, there is a physical distance value between two grid points, which is not convenient for direct subsequent processing. Therefore, a grid transformation strategy is used to convert the non-uniform grid in the three-dimensional physical space into a uniform grid in the three-dimensional computational space. Each grid point has relevant position information and other relevant field information in the Cartesian coordinate system. Then the grid boundary is processed. During training, the input is the local flow field block of each point, but the points near the boundary do not have enough adjacent blocks to obtain local patches, so it is necessary to fill the grid with the properties of the original flow field to obtain the preprocessed data; Process the grid boundary and fill the grid according to the original flow field characteristics to obtain optimized flow field data; (3) Divide the training data set and the test data set. First, use 80% of the data as the training data set and the remaining 20% ​​of the data as the test data set. 80% of the training data set is used as the training set and the remaining 20% ​​is used as the validation set. The IVD label calculated in step (1) is used as the training data. The vortex structure label is identified using the IVD (instantaneous vorticity deviation)-based vortex structure recognition method. (4) Normalize and standardize the flow field data. First, use the Z-score equation to standardize the entire vorticity field, and then use the sigmoid function to normalize the vorticity field between 0 and 1, so that the normalized vorticity field is obtained; (5) Constructing a network model. The present invention uses U-Net as a basic model and designs a 3D U-Net for vortex structure recognition, called MFV-Net. Three multi-scale feature enhancement modules are designed to help the network fully combine the global context information of the flow field and improve the recognition accuracy of the vortex structure. A 3D U-Net network model named MFV-Net is designed and constructed, with three built-in multi-scale vortex feature information enhancement modules, aiming to improve the network's ability to segment vortex structures. First, to solve the problem of gradient disappearance, an improved residual module (Res-Block) is used to replace the convolution block of U-Net. Secondly, to reduce information loss in the encoding and decoding process, a multi-scale connection enhancement module (MCE) is constructed at the jump connection of U-Net. Finally, to capture multi-scale feature information and improve the network's ability to recognize vortex structures, a multi-scale feature fusion module (MSFF) is introduced at the bottom of U-Net. (6) Design a loss function. In a physical flow field, points close to the boundary of an object shape are more important than other points, and the importance of a point is inversely proportional to the distance. In order to retain the characteristics of the physical flow field, a weighted cross entropy loss function is used. A weighted loss function related to the boundary position of the flow field is formulated to optimize the training process. (7) (7) Flow field blocks are sampled from the training data set and the network model is trained. The input is the coordinate information of the flow field, the three components of velocity (U, V, W), pressure, vorticity, instantaneous vorticity deviation, and weight ω. i , randomly sample flow field blocks of size 32×32×32 from the training dataset as the input of MFV-Net; Extract flow field block samples from the training data set, and the input includes flow field coordinates, velocity components (U, V, W), pressure, vorticity, instantaneous vorticity deviation and weight ω i ; (8) Post-processing: Reconstruct the combined flow field based on the position coordinate information of each flow field block output by the network, and finally obtain the vortex structure prediction result of the entire flow field.

2. In step (5) of claim 1, it is characterized in that: A U-Net network is integrated and designed for vortex structure recognition in flow field, called MFV-Net; according to requirement 1, a new vortex recognition network MFV-Net based on U-Net is designed. In step (5), a 3D U-Net is designed for vortex structure recognition based on U-Net as the network framework, called MFV-Net. Three multi-scale feature enhancement modules are designed to help the network fully combine the global context information of the flow field and improve the accuracy of vortex structure recognition. First, in order to solve the gradient disappearance problem, the improved residual module (Res-Block) is used to replace the convolution block of U-Net. The directional enhanced residual module (OER-Block) is used to replace the original convolution block; secondly, in order to reduce the information loss in the encoding and decoding process, a multi-channel feature enhancement module (MCFE) and a multi-scale connection enhancement module (MCE) are constructed at the encoder-decoder jump connection of MFV-NetU-Net; finally, in order to capture multi-scale feature information and improve the network's recognition ability of vortex structure, a multi-scale feature fusion and reconstruction module (MSR) is introduced at the bottom of MFV-NetU-Net. Multi-scale feature fusion module (MSFF); The MFV-Net structure is shown in Figure 1, which consists of a contraction and expansion part. Each step contains two Res-Block convolution operations. In the last layer of each step, a 2×2×2 maximum pooling layer is used for downsampling to reduce the size of the feature map. After each downsampling step, the number of feature channels doubles. The expansion part includes a 2×2×2 deconvolution (DeConv) operation to reduce the number of channels by half, and then it is concatenated with the corresponding feature map of the encoder after MCE feature enhancement. Finally, the same Res-Block convolution operation as the contraction part is applied. In order to overcome the overfitting problem, a dropout layer is added after the last few convolution layers. In the last layer of the network, a 1×1×1 convolution is used to reduce the number of output channels to achieve a binary classification operation for dividing the flow field vortex points and non-vortex points, thereby accurately identifying the flow field vortex structure.

3. In step (5) described in claim 1, the convolution block of the original U-Net is replaced by an improved residual module (Res-Block), and the residual network is applied to solve the performance degradation problem that occurs as the network deepens. The improved residual module is used to replace the conventional convolution block in U-Net in the encoding and decoding paths, which can enhance the network's vortex structure feature learning ability. The improved residual module is shown in Figure 2, where the parameterized rectified linear unit (PReLU) function is a variant of the classic rectified linear unit activation function (ReLU). Using PReLU to replace ReLU in the residual module can make the network more robust and increase the network's ability to resist noise interference. Its expression is as follows: The core formula of the residual module is: X l+1 =h(X l )+F(X l ,W l ) In the formula: First, h(Xl) represents the identity mapping of the input. In this process, a 1×1 convolution is used to adjust the channel dimension to ensure that the feature information can be directly transmitted between the low layer and the high layer, thereby ensuring the stability of the training gradient. F(Xl, Wl) represents the residual mapping part, which consists of two 3×3 convolution layers. After the convolution layer, a batch normalization layer (BN) and a PReLU activation function are added. The batch normalization layer can keep the distribution of each batch of data consistent, thereby accelerating the convergence of the network. In step (5) described in requirement 1, it is characterized in that: The directional enhanced residual module (OER-Block) is designed to replace the original convolution block. The OER-Block captures the local directional features in the flow field through a directional filter and adopts a multi-scale feature fusion strategy to capture vortex features of multiple scales from local details to global structures. The implementation of direction-aware convolution can be achieved by defining a directional convolution kernel G(θ) to extract features in a specific direction. The direction-enhanced convolution output Y of the convolution operation for the input feature map X can be expressed as: Y=G(θ)*X.

4. In step (5) described in claim 1, a multi-scale connection enhancement module (MCE) is constructed at the jump connection of U-Net. The MCE is located at the jump connection corresponding to the position of the codec. By adding the vortex structure characteristic information in the coding layer corresponding to the current decoding layer, efficient transmission of the coding result is achieved. The MCE structure is shown in Figure 3. The MCE assists in enhancing the flow field vortex structure characteristic information in the coding layer corresponding to the current decoding layer, thereby achieving efficient transmission of the coding result. Its mathematical definition is as follows: P1=MCE(X1,upsample(X2),upsample(upsample(X3)))) P2=MCE(X2,downsample(X1),upsample(X3)) P3=MCE(X3,downsample(X2),downsample(downsample(X1))) in, Upsample represents the upsampling operation using transposed convolution, downsample represents the downsampling operation using maximum pooling, X1, X2, and X3 represent the encoder outputs at different stages, respectively, and P1, P2, and P3 represent the operation results after the multi-scale connection enhancement module. First, in order to keep the dimensions of the output data of different coding layers consistent, maximum pooling is used for downsampling or transposed convolution is used for upsampling. In order to make full use of the coding outputs at different levels, the coding outputs except the X1 level are directly fused, and a 3×3 convolution layer is used to extract the spatial information of the fusion result. Then, in order to generate the channel attention mechanism, the 3×3 convolution layer is connected to the average pooling layer (avgpool) and the maximum pooling layer (maxpool), and the Softmax function is used to generate the channel attention feature map. Finally, a residual connection is used to fuse it with the generated output feature map. By using MCE, the information loss in the encoding and decoding process is reduced and the vortex structure characteristics are enhanced. In step (5) described in claim 1, it is characterized in that: a multi-channel feature enhancement module (MCFE) is designed, and the MCFE integrates vortex feature information at different levels through deformable convolution and inter-layer feature interaction mechanism; The structure of MCFE is shown in Figure 3, and its mathematical definition is as follows: P i =MCFE(X i ,concat(DeformConv(X i-1 ), upsample(Y i+1 )))Wherein, DeformConv represents the deformable convolution operation, concat represents the concatenation of feature maps of different channels, upsample represents upsampling using transposed convolution, Xi, Yi represent the codec outputs of different stages, and Pi represents the operation result after MCFE.

5. In step (5) described in claim 1, a multi-scale feature fusion module (MSFF) is designed and added to the bottom layer of U-Net. The module is a variant of the atrous spatial convolutional pooling pyramid (ASPP). The structure diagram of the MSFF module is shown in FIG4 ; In step (5) described in claim 1, it is characterized in that: A multi-scale feature fusion and reconstruction module (MSR) is designed. MSR extracts vortex features from different scales through a hierarchical network. After feature extraction, a learnable weight ωi is assigned to the feature map of each scale. These weights are dynamically adjusted during the training process and calculated through feature statistics. Finally, the weighted feature maps are summed to form the final fused feature map Ffinal. Among them, Fi is the i-th feature map, ωi is the corresponding weight; In the fused feature map, the weights between channels are generated to dynamically adjust the expression of features: F adjusted =F final ·σ(W·AvgPool(F final )) Among them, σ is the activation function, and W is the weight obtained by the fully connected layer; Finally, the deconvolution layer is used to restore the small-scale vortex details in the fused feature map. This step ensures that the information of the small-scale vortex structure is effectively preserved in multiple convolution operations.

Images