CN115795683A - Wing profile optimization method fusing CNN and Swin transform network - Google Patents
Wing profile optimization method fusing CNN and Swin transform network Download PDFInfo
- Publication number
- CN115795683A CN115795683A CN202211572055.0A CN202211572055A CN115795683A CN 115795683 A CN115795683 A CN 115795683A CN 202211572055 A CN202211572055 A CN 202211572055A CN 115795683 A CN115795683 A CN 115795683A
- Authority
- CN
- China
- Prior art keywords
- flow field
- swin
- airfoil
- cnn
- conv
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses an airfoil profile optimization method fusing CNN and Swin transducer networks, which comprises the following specific steps: characterizing the airfoil shape in a parameterization mode, acquiring a real flow field solution value in a computational fluid dynamics numerical simulation mode, and constructing a required data set; carrying out normalization pretreatment on the obtained airfoil profile parameters and the real flow field data set; carrying out linear coding on the coordinates of the parameterized control points of the profile shape, and extending and filling characteristic dimensions; reshaping the feature maps of adjacent dimensions into a high-resolution feature map by using a bilinear upsampling mode; constructing a Swin Conv Transformer Block basic network module by using a CNN network and a Swin Transformer to extract flow field characteristics; performing iterative calculation, and comparing the network prediction flow field with the real flow field value until the maximum training step length is reached; acquiring a flow field image of a corresponding airfoil profile by using the trained optimal model; and determining the optimized direction of the airfoil profile through the comparison of the flow field characteristics of the series of airfoil profiles.
Description
Technical Field
The invention belongs to the technical field of airfoil optimization, and particularly relates to an airfoil optimization method fusing CNN and Swin transform networks.
Background
The design process of the industrial airplane is a complex process, and the overall aerodynamic performance of the wing profile is determined to a certain extent by the optimized design of the wing profile; the flow field visually reflects the overall aerodynamic performance, so the flow field is usually considered in airfoil design and optimization.
The airfoil optimized design usually requires numerical simulation using computational fluid mechanics, but the process usually takes a large amount of computation time because the design space of the airfoil profile often has features with high dimensionality. In order to solve the problem, a related learner performs dimension reduction processing on a design space, for example, in the traditional method, a low-dimension parameter equation is constructed, and then related parameters are optimized; a common response surface model, a kriging model, and the like are widely used. However, these methods are limited to obtaining the aerodynamic parameters, and under the optimization constraint condition, the purpose of obtaining the maximum lift-drag ratio is taken as the optimization. However, this also leads to the problem that evaluating airfoil optimization solely with aerodynamic parameters is incomplete, it discards detailed aerodynamic features in the flow field, and this optimization is more empirical than strictly theoretical and thus also one-sided for airfoil optimization. A strict relation between a flow field structure and aerodynamic force is established in modern aerodynamic theory, and experienced researchers can determine the optimized direction of the airfoil profile by observing the flow field, so that the optimization method based on flow field visualization can greatly reduce trial and error cost of the traditional method, and quickly obtain the optimized airfoil profile meeting the target. The method and the device convert the traditional flow field prediction task into computer vision, and the important research is how to efficiently and accurately obtain the flow field characteristics directly.
Disclosure of Invention
The invention aims to solve the technical problem of providing an airfoil profile optimization method fusing CNN and Swin transducer networks, which takes an airfoil profile parameterized control point as an input label, quickly outputs a corresponding flow field image, realizes guidance of airfoil profile optimization by acquiring the change of serial airfoil profile flow field characteristics, and greatly reduces the time magnitude of an operation process while obtaining high precision and high accuracy.
In order to solve the technical problem, the invention is realized by the following modes:
an airfoil profile optimization method fusing a CNN and a Swin transform network comprises the following specific steps:
(1) Characterizing the airfoil shape in a parameterization mode, acquiring a real flow field solution value in a computational fluid dynamics numerical simulation mode, and constructing a required data set;
(2) Carrying out normalization pretreatment on the obtained airfoil profile parameters and the real flow field data set;
(3) Carrying out linear coding on the coordinates of the parameterized control points of the profile shape, and extending and filling characteristic dimensions;
(4) Reshaping the feature maps of adjacent dimensions into a high-resolution feature map by using a bilinear upsampling mode;
(5) Constructing a Swin Conv Transformer Block basic network module by using a CNN network and a Swin Transformer to extract flow field characteristics;
(6) Performing iterative calculation, and comparing the network prediction flow field with the real flow field value until the model converges or the maximum training step length is reached;
(7) Acquiring a flow field image of a corresponding wing profile by using the trained model;
(8) And determining the optimized direction of the airfoil profile through the comparison of the flow field characteristics of the series of airfoil profiles.
The method comprises the following steps of (1) representing the airfoil shape in a parameterization mode, acquiring a real flow field solution value in a computational fluid dynamics numerical simulation mode, and constructing a required data set, wherein the method specifically comprises the following steps:
(11) Setting control parameters as NURBS curve control point coordinates, and generating a series of airfoil shape control point coordinates by adopting a Latin hypercube sampling method;
(12) Generating an airfoil B-spline curve by using the corresponding control point coordinate parameters in the step (11);
(13) Dividing the grid of the airfoil external flow field structure by using grid generation software;
(14) Establishing a Navier-Stocks equation model, importing a grid to solve the airfoil real flow field, and obtaining a real flow field data set.
Further, the preprocessing the acquired data set in the step (2) specifically includes the following steps:
(21) Normalizing each airfoil profile parameter and the resolved real flow field data set, wherein the specific normalization mode is shown as a formula (1):
wherein X represents the normalized value, X represents the original value of the data set, and X min Denotes the minimum value, x, in the data set max Represents the maximum value in the dataset;
(22) And (4) cutting the data set subjected to the normalization processing in the step (21) into an 80% training set and a 20% testing set, wherein the testing set is used for verifying the final model prediction effect.
Further, the bilinear upsampling method in the step (4) specifically includes the following steps:
the bilinear upsampling mode is a mixed mode combining bilinear interpolation and pixel scrambling, and 2 times of amplification is carried out on the characteristic resolution ratio under the condition of keeping the characteristic dimension unchanged; bilinear interpolation reduces the difference in magnitude between the interpolated and original input values, and more detailed features are sampled by pixel scrambling operations.
In the step (5), a Swin Conv transform Block basic network module is constructed by using a CNN network and a Swin transform to extract the flow field characteristics, and the specific steps are as follows:
each basic network module consists of a specification layer, a multi-head self-attention module and a Conv layer, wherein BatchNorm (BN) and Relu activation functions are used, and the multi-head self-attention module comprises a Conv multi-head self-attention (CW-MSA) module and a Conv shift window-based multi-head self-attention (CSW-MSA) module which are respectively positioned in two continuous SCTB modules; by the shift window partitioning method, the continuous Swin Conv Transformer Block (SCTB) is expressed as:
wherein
And z l Respectively representing the output characteristics of the CW-MSA block and the Conv block of the l-th block,
and z l+1 Respectively representing the output characteristics of the CSW-MSA module and the Conv module of the l +1 th block; BN represents performing a BatchNorm normalization operation, conv represents performing a convolution operation, CW-MSA represents a Conv multi-headed self-attention operation, CSW-MSA represents a multi-headed self-attention block of a shift window;
the calculation formula of the specific attention mechanism is as follows:
wherein Q, K and V respectively represent a query, key and value matrix, d represents an input dimension, B represents a bias matrix, and Softmax represents an activation function.
In the step (6), iterative computation is performed, and the network prediction flow field value is continuously compared with the real flow field value until the model reaches the maximum training step length, and the method specifically comprises the following steps:
in the model training process, adam is used as an optimizer, normalization operation is carried out in a Batch Norm mode, an activation function is set to be RELU, a loss function is MSE, specific model training is carried out, meanwhile, in the training process, a Batch parameter Batch size is set to be 40, and the maximum training step size is 1000; and storing the optimal model in the training process, and finally outputting the flow field image through the CNN layer by the network.
Compared with the prior art, the invention has the following beneficial effects:
the method predicts the flow field of the airfoil by a deep learning framework fusing CNN and Swin transform and inputting the parameterized control points related to the airfoil shape; dimension amplification is carried out by adopting a double up-sampling mode, so that the chessboard effect of the traditional transposition convolution is solved; based on the combined architecture design of CNN and Swin transducer under deep learning, the extraction capability of the flow field is fully extracted, the dependence relationship of the global space of the flow field is learned by using the self-attention mechanism of Swin transducer, and the local flow field features are extracted by using CNN, so that the feature extraction capability of the network is improved, the time cost is reduced, and higher prediction accuracy is achieved; meanwhile, a rapid and accurate flow field prediction model is utilized to guide the airfoil optimization direction.
Drawings
FIG. 1 is a block flow diagram of the present invention;
FIG. 2 is a schematic diagram of the overall network architecture of the present invention;
FIG. 3 is a schematic diagram of a dual upsampling module structure according to the present invention;
FIG. 4 is a schematic diagram of a Swin Conv Transformer Block network according to the present invention;
FIG. 5 is a schematic diagram of the convolution attention mechanism of the present invention;
FIG. 6 is a schematic diagram comparing the model prediction with the real flow field according to the present invention;
FIG. 7 is a schematic diagram of the predicted results of different airfoil flow fields according to the present invention.
Detailed Description
The following detailed description of embodiments of the invention is provided in connection with the accompanying drawings and the examples.
As shown in fig. 1-2, an airfoil profile optimization method fusing CNN and Swin Transformer networks includes the following specific steps:
(1) Characterizing the airfoil shape in a parameterization mode, acquiring a real flow field solution value in a computational fluid dynamics numerical simulation mode, and constructing a required data set;
(2) Preprocessing the acquired data set: respectively carrying out normalization processing on the airfoil profile parameters and the real flow field data;
(3) Carrying out linear coding on the parameterized control point coordinates of the airfoil shape, taking the parameterized airfoil shape control point coordinates as input, carrying out linear coding by using a linear embedding layer, and extending and filling feature dimensions;
(4) Reshaping the feature maps of adjacent dimensions into a high-resolution feature map by using a bilinear upsampling mode;
(5) Constructing a Swin Conv Transformer Block basic network module by using a CNN network and a Swin Transformer to extract flow field characteristics;
(6) Performing iterative calculation, and comparing the network prediction flow field with the real flow field value until the model converges or the maximum training step length is reached;
(7) Acquiring a flow field image of a corresponding wing profile by using the trained model;
(8) And determining the optimized direction of the airfoil profile through the comparison of the flow field characteristics of the series of airfoil profiles.
The method comprises the following steps of (1) representing the airfoil shape in a parameterization mode, acquiring a real flow field solution value in a computational fluid dynamics numerical simulation mode, and constructing a required data set, wherein the method specifically comprises the following steps:
(11) Setting control parameters as NURBS curve control point coordinates, and generating a series of airfoil profile control point coordinates by adopting a Latin hypercube sampling method, namely airfoil profile parameters;
(12) Generating an airfoil B-spline curve by using the corresponding control point coordinate parameters in the step (11);
(13) Dividing the grid of the airfoil external flow field structure by using grid generation software;
(14) Establishing a Navier-Stocks equation model, importing a grid to solve the airfoil real flow field, and obtaining a real flow field data set.
The preprocessing of the acquired data set in the step (2) specifically comprises the following steps:
(21) Normalizing each airfoil profile parameter and the resolved real flow field data set, and mapping all data to a [0,1] interval in proportion, wherein the specific normalization mode is shown as a formula (1):
wherein X represents the normalized value, X represents the original value of the data set, and X min Represents the minimum value, x, in the data set max Represents the maximum value in the dataset;
(22) And (4) cutting the data set subjected to the normalization processing in the step (21) into 80% of training sets and 20% of test sets, wherein the test sets are used for verifying the final model prediction effect.
In the bilinear upsampling mode in the step (4), the specific steps are as follows:
as shown in fig. 3, the specific bilinear upsampling design structure has a bilinear upsampling mode that is a hybrid mode combining bilinear interpolation and pixel scrambling, and performs 2-fold amplification on the feature resolution while keeping the feature dimension unchanged; bilinear interpolation reduces the size difference between the interpolated and original input values, and more detailed features are sampled by pixel scrambling operations.
In the step (5), a Swin Conv transform Block basic network module is constructed by using a CNN network and a Swin transform to extract the flow field characteristics, and the specific steps are as follows:
the Swin Conv Transformer Block (SCTP) structure is shown in FIG. 4, each SCTP is composed of a specification layer, a multi-head self-attention module and a Conv layer, a Batch Norm (BN) and a Relu activation function are used in the SCTP, the Conv layer represents a CNN network, the whole model architecture is based on a Swin-Transformer module, the CNN network is fused into the Swin-Transformer module, and a new SCTP module is constructed; instead of Layer Norm (LN) and GELU in the traditional transform block, this change speeds up the model training and results in a slight improvement in accuracy.
The multi-head self-attention module comprises a Conv multi-head self-attention (CW-MSA) module and a Conv shift window-based multi-head self-attention (CSW-MSA) module, the two modules are respectively positioned in two continuous SCTB modules, one of the two continuous SCTB modules comprises a specification layer, a Conv multi-head self-attention (CW-MSA) module, a Conv layer and a specification layer, and the other multi-head SCTB module comprises a specification layer, a Conv shift window-based multi-head self-attention (CSW-MSA) module, a Conv layer and a specification layer; by the shift window partitioning method, the continuous Swin Conv Transformer Block (SCTB) is expressed as:
wherein
And z l Conv-modulus representing CW-MSA-module and l-th block, respectivelyThe output characteristics of the block or blocks are,
and z l+1 Respectively representing the output characteristics of the CSW-MSA module and the Conv module of the l +1 th block; BN represents performing a BatchNorm normalization operation, conv represents performing a convolution operation, CW-MSA represents a Conv multi-headed self-attention operation, and CSW-MSA represents a multi-headed self-attention block of the shift window.
The calculation formula of the specific attention mechanism is as follows:
wherein Q, K, V represent query, key and value matrices, respectively, d represents input dimensions, B represents a bias matrix, and Softmax represents an activation function.
Before performing attention mechanism calculation, feature extraction is performed on spatial and channel dimensions respectively by using convolution of 1x1 and convolution kernel of 3x3, feature extraction is performed by converting a multilayer perceptron layer of a traditional Transformer into a convolution layer, the SCTB layer number of each Block is respectively set to be 2,4, and the multi-head attention number of each Block is respectively 2, 4.
Iterative computation is carried out in the step (6), the network prediction flow field is compared with the real flow field value until the model converges or the maximum training step length is reached, and the specific steps are as follows:
in the model training process, adam is used as an optimizer, normalization operation is carried out in a Batch Norm mode, an activation function is set to Relu, a loss function is MSE, specific model training is carried out, meanwhile, in the training process, a Batch processing parameter Batch size is set to 40, the maximum training step size is 1000, the model with the minimum loss function in the training process is saved as an optimal model, and finally the flow field image is output through a CNN layer by a network.
Further, the loss function MSE is an expression of mean square path difference, and the formula is as follows:
wherein, y i Data representing the real flow field is then transmitted,
and the predicted value of the flow field is represented, m represents the number of data sets, and the smaller the value of MSE is, the better performance is represented.
As shown in fig. 6, the result of comparing the model prediction with the real flow field shows that the prediction result can well predict the real flow field characteristics, and the prediction result and the real flow field result are close to 0 in terms of absolute error distribution, which shows a good prediction effect.
In the step (8), the optimized direction of the airfoil is determined by observing the change of the structural characteristics of the tail vortex of the flow field structure of the series of airfoils. As shown in fig. 7, the prediction results of different airfoil flow field characteristics show that the extracted models can achieve good prediction results; by further observing the change of the flow field characteristics, the vortex structure at the tail part of the Airfoil 3 is obviously weakened compared with that of the Airfoil 1 Airfoil, so that the aerodynamic performance is improved, and the optimized Airfoil with better aerodynamic performance than that of the Airfoil 1 is obtained.
The foregoing is illustrative of the present invention and it will be appreciated that modifications and variations may be made thereto without departing from the spirit and scope of the invention as defined by the appended claims.
Claims (6)
1. A wing profile optimization method fusing CNN and Swin Transformer networks is characterized by comprising the following steps: the method comprises the following specific steps:
(1) Characterizing the airfoil shape in a parameterization mode, acquiring a real flow field solution value in a computational fluid dynamics numerical simulation mode, and constructing a required data set;
(2) Carrying out normalization pretreatment on the obtained airfoil profile parameters and the real flow field data set;
(3) Carrying out linear coding on the coordinates of the parameterized control points of the profile shape, and extending and filling characteristic dimensions;
(4) Reshaping the feature maps of adjacent dimensions into a high-resolution feature map by using a bilinear upsampling mode;
(5) Constructing a Swin Conv Transformer Block basic network module by using a CNN network and a Swin Transformer to extract flow field characteristics;
(6) Performing iterative calculation, and comparing the network prediction flow field with the real flow field value until the model converges or the maximum training step length is reached;
(7) Acquiring a flow field image of a corresponding airfoil profile by using the trained model;
(8) And determining the optimized direction of the airfoil profile through comparing the flow field characteristics of the series airfoil profiles.
2. The method for optimizing an airfoil profile by fusing a CNN and a Swin transducer network as claimed in claim 1, wherein:
in the step (1), the airfoil profile is characterized in a parameterization mode, a computational fluid dynamics numerical simulation mode is adopted to obtain a real flow field solution value, and a required data set is constructed, and the method specifically comprises the following steps:
(11) Setting control parameters as NURBS curve control point coordinates, and generating a series of airfoil profile shape control point coordinates by adopting a Latin hypercube sampling method;
(12) Generating an airfoil profile B spline curve by using the corresponding control point coordinate parameters in the step (11);
(13) Dividing the grid of the airfoil external flow field structure by using grid generation software;
(14) Establishing a Navier-Stocks equation model, importing a grid to solve the airfoil real flow field, and obtaining a real flow field data set.
3. The method for optimizing an airfoil profile by fusing a CNN and a Swin transform network as claimed in claim 1, wherein:
the preprocessing of the acquired data set in the step (2) specifically comprises the following steps:
(21) Normalizing each airfoil profile parameter and the resolved real flow field data set, wherein the specific normalization mode is shown as a formula (1):
wherein X represents the normalized value, X represents the original value of the data set, and X min Represents the minimum value, x, in the data set max Represents the maximum value in the dataset;
(22) And (4) cutting the data set subjected to the normalization processing in the step (21) into an 80% training set and a 20% testing set, wherein the testing set is used for verifying the final model prediction effect.
4. The method for optimizing an airfoil profile by fusing a CNN and a Swin transducer network as claimed in claim 1, wherein:
the bilinear upsampling mode in the step (4) comprises the following specific steps:
the bilinear upsampling mode is a mixed mode combining bilinear interpolation and pixel scrambling, and under the condition of keeping the feature dimension unchanged, the feature resolution is amplified by 2 times; bilinear interpolation reduces the size difference between the interpolated and original input values, and more detailed features are sampled by pixel scrambling operations.
5. The method for optimizing an airfoil profile by fusing a CNN and a Swin transducer network as claimed in claim 1, wherein:
in the step (5), a Swin Conv transform Block basic network module is constructed by using a CNN network and a Swin transform to extract flow field characteristics, and the specific steps are as follows:
each basic network module consists of a specification layer, a multi-head self-attention module and a Conv layer, wherein a Batch Norm and Relu activating function is used in the basic network module, the multi-head self-attention module comprises a Conv multi-head self-attention module and a Conv shift window-based multi-head self-attention module, and the two modules are respectively positioned in two continuous SCTB modules; by the shift window partitioning method, the continuous Swin Conv Transformer Block is expressed as:
wherein
And z l Respectively representing the output characteristics of the CW-MSA block and the Conv block of the l-th block,
and z l+1 Respectively representing the output characteristics of the CSW-MSA module and the Conv module of the l +1 th block; BN represents performing a BatchNorm normalization operation, conv represents performing a convolution operation, CW-MSA represents a Conv multi-headed self-attention operation, CSW-MSA represents a multi-headed self-attention block of a shift window;
the calculation formula of the specific attention mechanism is as follows:
wherein Q, K and V respectively represent a query, key and value matrix, d represents an input dimension, B represents a bias matrix, and Softmax represents an activation function.
6. The method for optimizing an airfoil profile by fusing a CNN and a Swin transform network as claimed in claim 1, wherein:
in the step (6), iterative computation is performed, and the network prediction flow field is compared with the real flow field value until the model converges or the maximum training step length is reached, and the specific steps are as follows:
in the model training process, adam is used as an optimizer, normalization operation is carried out in a Batch Norm mode, an activation function is set to Relu, a loss function is MSE, specific model training is carried out, meanwhile, in the training process, a Batch processing parameter Batch size is set to 40, the maximum training step size is 1000, the optimal model in the training process is stored, and finally the flow field image is output through a CNN layer by a network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211572055.0A CN115795683B (en) | 2022-12-08 | 2022-12-08 | Airfoil optimization method integrating CNN and Swin converter network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211572055.0A CN115795683B (en) | 2022-12-08 | 2022-12-08 | Airfoil optimization method integrating CNN and Swin converter network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115795683A true CN115795683A (en) | 2023-03-14 |
| CN115795683B CN115795683B (en) | 2023-07-21 |
Family
ID=85419058
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211572055.0A Active CN115795683B (en) | 2022-12-08 | 2022-12-08 | Airfoil optimization method integrating CNN and Swin converter network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115795683B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116363603A (en) * | 2023-03-20 | 2023-06-30 | 哈尔滨市科佳通用机电股份有限公司 | Deep learning-based car ladder handrail detection method, storage medium and equipment |
| CN117272806A (en) * | 2023-09-22 | 2023-12-22 | 四川大学 | Airfoil separation flow prediction method based on flow field inversion and machine learning |
| CN117421997A (en) * | 2023-12-15 | 2024-01-19 | 中国空气动力研究与发展中心计算空气动力研究所 | Method and device for determining wing-shaped flow field information, terminal equipment and storage medium |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120285648A1 (en) * | 2011-05-10 | 2012-11-15 | Howmet Corporation | Ceramic core with composite insert for casting airfoils |
| CN111295657A (en) * | 2017-09-08 | 2020-06-16 | 洛桑联邦理工学院(Epfl) | Shape optimization via gradient descent technology device using convolutional neural network proxy |
| CN113849910A (en) * | 2021-09-23 | 2021-12-28 | 四川大学 | Dropout-based BiLSTM network wing resistance coefficient prediction method |
| CN114065662A (en) * | 2021-11-12 | 2022-02-18 | 西北工业大学 | Method suitable for rapidly predicting airfoil flow field with variable grid topology |
| CN114399646A (en) * | 2021-12-21 | 2022-04-26 | 北京中科明彦科技有限公司 | Image description method and device based on Transformer structure |
| CN114494254A (en) * | 2022-04-14 | 2022-05-13 | 科大智能物联技术股份有限公司 | Product appearance defect classification method based on fusion of GLCM and CNN-Transformer and storage medium |
| CN115082293A (en) * | 2022-06-10 | 2022-09-20 | 南京理工大学 | Image registration method based on Swin transducer and CNN double-branch coupling |
| CN115100490A (en) * | 2022-07-20 | 2022-09-23 | 广东工业大学 | Image quality evaluation method and system and readable storage medium |
| CN115294073A (en) * | 2022-08-11 | 2022-11-04 | 西安电子科技大学广州研究院 | Texture object defect detection method based on knowledge distillation |
| CN115293050A (en) * | 2022-08-22 | 2022-11-04 | 上海大学 | Method, device and system for establishing fluid flow reduced-order model and storage medium |
-
2022
- 2022-12-08 CN CN202211572055.0A patent/CN115795683B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20120285648A1 (en) * | 2011-05-10 | 2012-11-15 | Howmet Corporation | Ceramic core with composite insert for casting airfoils |
| CN111295657A (en) * | 2017-09-08 | 2020-06-16 | 洛桑联邦理工学院(Epfl) | Shape optimization via gradient descent technology device using convolutional neural network proxy |
| CN113849910A (en) * | 2021-09-23 | 2021-12-28 | 四川大学 | Dropout-based BiLSTM network wing resistance coefficient prediction method |
| CN114065662A (en) * | 2021-11-12 | 2022-02-18 | 西北工业大学 | Method suitable for rapidly predicting airfoil flow field with variable grid topology |
| CN114399646A (en) * | 2021-12-21 | 2022-04-26 | 北京中科明彦科技有限公司 | Image description method and device based on Transformer structure |
| CN114494254A (en) * | 2022-04-14 | 2022-05-13 | 科大智能物联技术股份有限公司 | Product appearance defect classification method based on fusion of GLCM and CNN-Transformer and storage medium |
| CN115082293A (en) * | 2022-06-10 | 2022-09-20 | 南京理工大学 | Image registration method based on Swin transducer and CNN double-branch coupling |
| CN115100490A (en) * | 2022-07-20 | 2022-09-23 | 广东工业大学 | Image quality evaluation method and system and readable storage medium |
| CN115294073A (en) * | 2022-08-11 | 2022-11-04 | 西安电子科技大学广州研究院 | Texture object defect detection method based on knowledge distillation |
| CN115293050A (en) * | 2022-08-22 | 2022-11-04 | 上海大学 | Method, device and system for establishing fluid flow reduced-order model and storage medium |
Non-Patent Citations (5)
| Title |
|---|
| LIANG HUANG等: "Fine-Grained Ship Classification by Combining CNN and Swin Transformer", REMOTE SENSING, pages 1 - 7 * |
| ZE LIU等: "Swin Transformer: Hierarchical Vision Transformer using Shifted Windows", COMPUTER VISION AND PATTERN RECOGNITION, pages 1 - 15 * |
| ZHULIANG YAO等: "Leveraging Batch Normalization for Vision Transformers", 2021 IEEE/CVF INTERNATIONAL CONFERENCE ON COMPUTER VISION WORKSHOPS, pages 1 - 10 * |
| 占哲琦;陈鹏;桑永胜;彭德中;: "融合双注意力的深度神经网络在无人机目标检测中的应用", 现代计算机, no. 11, pages 31 - 36 * |
| 张玄武;郑耀;杨波威;张继发;: "基于级联前向网络的翼型优化设计", 浙江大学学报(工学版), no. 07, pages 142 - 148 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116363603A (en) * | 2023-03-20 | 2023-06-30 | 哈尔滨市科佳通用机电股份有限公司 | Deep learning-based car ladder handrail detection method, storage medium and equipment |
| CN117272806A (en) * | 2023-09-22 | 2023-12-22 | 四川大学 | Airfoil separation flow prediction method based on flow field inversion and machine learning |
| CN117421997A (en) * | 2023-12-15 | 2024-01-19 | 中国空气动力研究与发展中心计算空气动力研究所 | Method and device for determining wing-shaped flow field information, terminal equipment and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115795683B (en) | 2023-07-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115795683A (en) | Wing profile optimization method fusing CNN and Swin transform network | |
| CN110570396B (en) | Industrial product defect detection method based on deep learning | |
| CN109165664B (en) | Attribute-missing data set completion and prediction method based on generation of countermeasure network | |
| WO2021000556A1 (en) | Method and system for predicting remaining useful life of industrial equipment, and electronic device | |
| Duru et al. | A deep learning approach for the transonic flow field predictions around airfoils | |
| CN110222828B (en) | Unsteady flow field prediction method based on hybrid deep neural network | |
| CN112101480A (en) | Multivariate clustering and fused time sequence combined prediction method | |
| CN112185104B (en) | Traffic big data restoration method based on countermeasure autoencoder | |
| CN114235330B (en) | Multi-source pneumatic load model construction method for correlation wind tunnel test and calculation data | |
| CN113627093B (en) | Underwater mechanism trans-scale flow field characteristic prediction method based on improved Unet network | |
| CN106407567B (en) | A kind of RV Parametric Design of Reducer modeling method | |
| CN114722686B (en) | Method for designing and optimizing lifting lug of large equipment based on finite element analysis | |
| CN113988449A (en) | Wind power prediction method based on Transformer model | |
| CN113094860A (en) | Industrial control network flow modeling method based on attention mechanism | |
| CN114692488A (en) | Method and device for generating supercritical airfoil, electronic equipment and storage medium | |
| CN115423796A (en) | Chip defect detection method and system based on TensorRT accelerated reasoning | |
| CN110598326A (en) | Well testing interpretation method based on artificial intelligence | |
| CN116894180B (en) | Product manufacturing quality prediction method based on different composition attention network | |
| CN116933141B (en) | Multispectral laser radar point cloud classification method based on multicore graph learning | |
| CN106777616A (en) | A kind of high-speed rice pot seedling planting machine seedling box mover Multidisciplinary Optimization method | |
| CN116958553A (en) | Lightweight plant point cloud segmentation method based on non-parametric attention and point-level convolution | |
| CN115206455B (en) | Deep neural network-based rare earth element component content prediction method and system | |
| CN111783335B (en) | Transfer learning-based few-sample structure frequency response dynamic model correction method | |
| CN114295967A (en) | Analog circuit fault diagnosis method based on migration neural network | |
| CN115496824B (en) | Multi-class object-level natural image generation method based on hand drawing |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |