A flow field image driven method for steady and unsteady aerodynamic modeling of airfoil

Abstract

The present invention belongs to the field of computational fluid dynamics and artificial intelligence, and is specifically a flow field image driven airfoil steady and unsteady aerodynamic modeling method, firstly, a sample airfoil group is selected and a sample data set is constructed, wherein the sample data includes a signed distance field image of the sample airfoil, a velocity cloud map under several steady states, and a synthetic image of the angle of attack-Mach number corresponding to the steady state, a sequence of velocity cloud maps under an unsteady state, and a synthetic image of the angle of attack-Mach number corresponding to the unsteady state; secondly, a U-Net neural network for steady flow field images and a U-Net-LSTM neural network for unsteady flow field images are established and trained; finally, the trained U-Net neural network and the U-Net-LSTM neural network are used to quickly predict the steady and unsteady flow fields of the predicted airfoil. The present invention can avoid airfoil geometry parameterization by constructing a training framework based on pixel features, and can simultaneously and efficiently predict the airfoil flow fields in steady and unsteady states.

Description

Technical Field

[0001] The invention belongs to the fields of computational fluid dynamics and artificial intelligence, and in particular relates to a flow field image driven airfoil steady and unsteady aerodynamic modeling method. Background Art

[0002] During the airfoil design process, designers need to iterate the airfoil multiple times and quickly evaluate various solutions, which requires obtaining the airfoil flow field and aerodynamic force multiple times. The traditional methods for obtaining the airfoil flow field and aerodynamic force are wind tunnel tests and CFD calculations, but wind tunnel tests are expensive; although CFD calculations do not require large-scale experimental equipment, repeated calls to CFD analysis during the design process make this method extremely inefficient, especially when calculating unsteady flow fields and aerodynamic forces, which requires more computers.

[0003] In fact, there is a certain mapping relationship between the flight state and the airfoil flow field and aerodynamic force. Therefore, in the design process, existing wind tunnel test data or CFD data are often used to model the airfoil aerodynamic force and establish a mapping relationship to achieve rapid prediction of the airfoil flow field and aerodynamic force. For steady aerodynamic modeling, polynomial fitting methods and spline interpolation methods can be used, such as angle of attack-aerodynamic force / moment ( ) data to interpolate and fit, and predict the aerodynamic data under unknown angle of attack; for unsteady aerodynamic modeling, time series analysis methods and dynamic mode decomposition methods can be used, for example, to use time-angle of attack-aerodynamic force / torque ( ) data to interpolate and fit, and then predict the aerodynamic data at unknown times and unknown states. However, the above traditional methods need to establish complex physical models and mathematical models according to specific problems, and the parameter tuning and training process of the model are complicated, time-consuming and labor-intensive. Especially for unsteady aerodynamic modeling, it is necessary to process time series data and dynamic characteristics, which further increases the complexity of the model and the difficulty of training; the traditional interpolation fitting method is highly dependent on training data, and the model has limited generalization ability under different flow field conditions and airfoil shapes, and it is difficult to adapt to a variety of application scenarios, resulting in unstable performance in practical applications; and although the traditional method has high prediction accuracy in theory, it is inefficient in actual calculations and cannot meet the needs of real-time prediction and rapid response.

[0004] With the rise of computer artificial intelligence methods, artificial intelligence methods such as machine learning have begun to be applied to the rapid prediction of airfoil flow fields and aerodynamic forces. Currently, the following methods are common: based on convolutional neural networks, extract airfoil geometry information, and then feed the extracted geometry information, Reynolds number and angle of attack as input to the feedforward neural network to obtain an approximate model for predicting the flow field; based on the feedforward neural network, the geometric parameters of the airfoil and the physical coordinates of the flow field points are used as the input of the neural network, and the velocity and pressure at the corresponding physical coordinates are used as the output to establish a transonic airfoil flow field prediction model; based on the Transformer network architecture, extract the geometric parameters of the airfoil, and further establish the mapping relationship between the geometric parameters, Reynolds number, angle of attack, flow field coordinates, wall distance and airfoil velocity field and pressure field; based on the deep neural network model, use structured grids to predict the airfoil flow field for aerodynamic conditions of various angles of attack, Reynolds number and Mach number. In general, the input of the neural network in the current airfoil aerodynamic modeling is parameterized airfoil geometry information, and the flow field information is predicted point by point based on the physical coordinates. This method is rather cumbersome, and the extraction of airfoil geometry information requires the establishment of a separate neural network model, and the cost and timeliness of predicting the flow field information point by point are poor. In addition, the current intelligent modeling of airfoil aerodynamics is either only for steady conditions or only for unsteady conditions, which limits the application scenarios of the current intelligent neural network method. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention proposes a flow field image driven airfoil steady and unsteady aerodynamic modeling method, which is based on U-Net neural network and U-Net-LSTM neural network, and by constructing a training framework based on pixel features, the airfoil geometric shape is converted into an SDF image, and a synthetic image of angle of attack-Mach number is constructed, which is mixed with the flow field image for training, so that the unknown steady and unsteady airfoil flow field can be predicted; compared with the traditional aerodynamic modeling method, the method of the present invention is based on image-driven training architecture, which can avoid the construction of complex mathematical and physical models; in addition, neural network modeling is directly performed on the airfoil geometric image and the flow field image, which can avoid the airfoil geometric parameterization, and can directly predict the flow field characteristics to give a velocity cloud map; since there is no need to predict the flow field information point by point, it is more concise and efficient, with lower information cost and higher timeliness.

[0006] The technical solution provided by the present invention is:

[0007] The method for modeling steady and unsteady aerodynamic forces of an airfoil driven by a flow field image comprises the following steps:

[0008] Step 1: Select a sample airfoil group and construct a sample data set;

[0009] For a sample airfoil in the sample airfoil group:

[0010] Based on the CFD method, velocity cloud diagrams of the sample airfoil under several steady states and a sequence of velocity cloud diagrams under unsteady states are calculated; the steady state is composed of a set angle of attack and a set incoming flow Mach number; the unsteady state is composed of a set angle of attack variation law and a set incoming flow Mach number;

[0011] According to the geometric image of the sample airfoil, a signed distance field image of the sample airfoil is obtained;

[0012] According to the geometric image of the sample airfoil and a certain constant state, a synthetic image of the angle of attack-Mach number corresponding to the certain constant state is calculated;

[0013] According to the geometric image of the sample airfoil, a certain angle of attack and a set incoming flow Mach number condition under an unsteady state, a synthetic image of the angle of attack-Mach number corresponding to the angle of attack and the set incoming flow Mach number condition is calculated;

[0014] The sample data corresponding to the sample airfoil are: a signed distance field image of the sample airfoil, velocity cloud images of the sample airfoil under several steady states and a composite image of the angle of attack-Mach number corresponding to the steady state, a sequence of velocity cloud images of the sample airfoil under an unsteady state and a composite image of the angle of attack-Mach number corresponding to various angles of attack and set incoming flow Mach number conditions under the unsteady state;

[0015] Step 2: constructing a U-Net neural network for steady flow field images; the U-Net neural network is an improved neural network based on a standard U-Net network model, with an input layer, a first encoder, a first decoder and an output layer as component modules; the input data of the U-Net neural network are a signed distance field image of the airfoil and a composite image of the angle of attack-Mach number, and the output data are a velocity cloud map of the airfoil in a steady state;

[0016] Step 3: construct a U-Net-LSTM neural network for unsteady flow field images; the U-Net-LSTM neural network is a U-Net neural network established in step 2 with an LSTM module added, and is composed of an input layer, a second encoder, a second decoder, an LSTM module, and an output layer; the input data of the U-Net-LSTM neural network is a signed distance field image of the airfoil, and an image sequence composed of a plurality of synthetic images of angle of attack-Mach number, and the output data is a velocity cloud map of the airfoil at the next moment in an unsteady state; the image sequence is a sequence of equal time intervals;

[0017] Step 4: Based on the sample data set obtained in step 1, a steady sample set is determined, and the U-Net neural network obtained in step 2 is trained using the steady sample set; based on the sample data set obtained in step 1, an unsteady sample set is determined, and the U-Net-LSTM neural network obtained in step 3 is trained using the unsteady sample set;

[0018] The steady sample set refers to a sample data set consisting of a signed distance field image of a sample airfoil, velocity cloud images of the sample airfoil under several steady states, and a synthetic image of angle of attack-Mach number corresponding to the steady state;

[0019] The unsteady sample set refers to a sample data set consisting of a signed distance field image of a sample airfoil, a velocity cloud map sequence of the sample airfoil in an unsteady state, and a synthetic image of the angle of attack-Mach number corresponding to various angles of attack and set incoming flow Mach number conditions in an unsteady state;

[0020] Step 5: Rapidly predict the steady and unsteady flow fields of the airfoil to be predicted;

[0021] For steady flow prediction:

[0022] Obtain a signed distance field image of the airfoil to be predicted, input the signed distance field image of the airfoil to be predicted and a synthetic image of the angle of attack-Mach number under a set steady state into a trained U-Net neural network, obtain a velocity cloud map of the airfoil to be predicted under a set steady state, and realize steady flow field prediction of the airfoil to be predicted;

[0023] For unsteady flow prediction:

[0024] The signed distance field image of the airfoil to be predicted is obtained, and an image sequence consisting of the signed distance field image of the airfoil to be predicted and a synthetic image of the angle of attack-Mach number under a set unsteady state is input into a trained U-Net-LSTM neural network to obtain the velocity cloud map of the airfoil to be predicted at the next moment under the set unsteady state, thereby realizing the unsteady flow field prediction of the airfoil to be predicted.

[0025] Furthermore, if the type of the airfoil to be predicted is known, the sample airfoils in the sample airfoil group are consistent with the type of the airfoil to be predicted; if the airfoil to be predicted is an airfoil obtained by optimizing the basic airfoil, the sample airfoils in the sample airfoil group are consistent with the type of the basic airfoil. The types of airfoils include NACA four-digit airfoils, NACA five-digit airfoils, RAE airfoils, etc.

[0026] Furthermore, in step 1, the process of calculating the velocity cloud diagrams of the sample airfoil under several steady states and the velocity cloud diagram sequence under unsteady states based on the CFD method is as follows:

[0027] For the steady state, in the given angle of attack range [ , ] and the incoming flow Mach number range [ , ], randomly select M angles of attack and L incoming Mach numbers Based on the CFD method, the velocity cloud diagram of the sample airfoil under the conditions of set angle of attack and set incoming flow Mach number is calculated, and a total of Velocity contours; thus creating a velocity contour for a sample airfoil training cases, created for each of the N airfoils in the sample airfoil group training cases;

[0028] For the unsteady state, the pitch oscillation of the sample airfoil around the 1 / 4 chord length point is taken as the research object, and the change process of its angle of attack is:

[0029]

[0030] Where: is the average angle of attack, is the angle of attack oscillation amplitude, is the oscillation frequency, is the time; reduce the frequency , where C is the reference length, V is the set flight speed, and the set reduction frequency range [ , ], select P reduction frequencies at equal intervals within the set reduction frequency range Based on the CFD method, the velocity cloud diagram sequence of the sample airfoil at the corresponding reduction frequency is calculated, and a total of P velocity cloud diagram sequences are obtained; a total of N airfoils in the sample airfoil group are created A sequence of training cases.

[0031] Furthermore, in step 1, the process of obtaining the signed distance field (SDF) image of the sample airfoil according to the geometric image of the sample airfoil is as follows:

[0032] Create a zero level set to represent the point set on the airfoil surface, denoted as Z, as follows:

[0033]

[0034] in, Indicates the plane area where the sample airfoil geometry image is located The point on is a level set function if and only if When located on the airfoil geometry boundary , if and only if When inside the airfoil geometry boundary , if and only if Outside the airfoil geometry boundary ;

[0035] Then, based on the level set function, the signed distance function is calculated as follows:

[0036]

[0037] Point Set The point on for The symbol; according to The numerical value of is mapped to obtain the SDF color image.

[0038] Furthermore, in step 1, the process of calculating the composite image of angle of attack-Mach number is:

[0039] The pixel value in the synthetic image is recorded as , the calculation expression is as follows:

[0040]

[0041] in is the pixel value of the corresponding pixel point in the airfoil geometry image, is the maximum value of pixels in the airfoil geometry image; is the incoming flow Mach number corresponding to the synthetic image, is the maximum Mach number set.

[0042] Furthermore, the processing process of the U-Net neural network constructed in step 2 for the input data is:

[0043] First, the input layer receives the input data and then passes it to the first encoder;

[0044] The first encoder includes multiple convolutional layers and maximum pooling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, a maximum pooling layer is connected to perform a maximum pooling operation on the input feature map, which is then used as the input data of the next convolutional layer; the feature map obtained by the last convolutional layer of the first encoder is input to the first decoder;

[0045] The first decoder includes multiple convolutional layers and upsampling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, an upsampling layer is connected to perform a bilinear interpolation upsampling operation on the input feature map, and the feature map obtained after the upsampling operation is concatenated with the feature map before the corresponding maximum pooling operation in the first encoder as the input data of the next convolutional layer, so that the spatial dimension of the feature map is restored layer by layer until the feature map obtained by the last convolutional layer is output to the output layer;

[0046] The final output data is obtained through the output layer.

[0047] Furthermore, the U-Net-LSTM neural network constructed in step 3 introduces an LSTM module on the basis of the U-Net neural network constructed in step 2. Specifically, the LSTM module is connected between the convolution layer before the maximum pooling operation in the second encoder and the convolution layer after the upsampling operation in the second decoder, and the multi-scale time series features are transmitted through the LSTM module.

[0048] Furthermore, the processing process of the U-Net-LSTM neural network constructed in step 3 for the input data is:

[0049] First, the input layer receives the input data and then passes it to the second encoder;

[0050] The second encoder includes multiple convolutional layers and maximum pooling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, a maximum pooling layer is connected to perform a maximum pooling operation on the input feature map, which is then used as the input data of the next convolutional layer; the feature map obtained by the last convolutional layer of the second encoder is input to the second decoder;

[0051] The feature map output by the convolution layer before the maximum pooling operation in the second encoder is also input into the LSTM module, through which the feature map is processed in the time dimension to extract the feature map in the time series;

[0052] The second decoder includes multiple convolutional layers and upsampling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, an upsampling layer is connected to perform a bilinear interpolation upsampling operation on the input feature map, and the feature map obtained by the bilinear interpolation upsampling operation is concatenated with the feature map output by the corresponding LSTM module as the input data of the next convolutional layer, so that the spatial dimension of the feature map is restored layer by layer until the feature map obtained by the last convolutional layer is output to the output layer;

[0053] The final output data is obtained through the output layer.

[0054] Furthermore, the convolution operation performed by the convolution layer on the input data is a convolution operation with a padding operation. The padding operation ensures that the height and width of the feature map remain unchanged after each convolution, so that no additional cropping operation is required when the feature map before the maximum pooling operation is spliced ​​with the feature map after the upsampling operation. The padding operation is as follows:

[0055] Assuming the size of the input feature map of the convolution layer is H×W, the size of the convolution kernel is K×K, the number of padding is P, and the stride is S, the size of the output feature map is:

[0056] +1

[0057] +1

[0058] The value of the padding number P is calculated based on the requirement that the size of the output feature map remains the same as the size of the input feature map, thereby ensuring that the flow field image feature map information is not lost.

[0059] Furthermore, after the convolution layer performs the convolution operation on the input data, the batch normalization method is used to normalize the distribution of each feature channel in the feature map obtained by the convolution operation; the specific calculation formula of batch normalization is:

[0060]

[0061] in, Express Perform batch normalization operation. is the input data, is the batch mean, is the standard deviation of the batch, is the scaling parameter, is the translation parameter.

[0062] Furthermore, the specific interpolation formula of the upsampling layer for performing bilinear interpolation upsampling operation on the input data is as follows:

[0063]

[0064] in, Representing coordinates The pixel value after upsampling is is the coordinate in the feature map before upsampling The pixel values ​​of the four nearest neighboring pixels.

[0065] Beneficial Effects

[0066] The flow field image driven airfoil steady and unsteady aerodynamic modeling method proposed in the present invention is based on U-Net neural network and U-Net-LSTM neural network, and by constructing a training framework based on pixel features, the airfoil geometric shape is converted into an SDF image, and a synthetic image about angle of attack-Mach number is constructed, and mixed training is performed with the flow field image, so that unknown steady and unsteady airfoil flow fields can be predicted; compared with the traditional aerodynamic modeling method, the method of the present invention is based on image-driven training framework, which can avoid complex mathematical and physical model construction; in addition, neural network modeling is directly performed on airfoil geometric images and flow field images, which can avoid airfoil geometric parameterization, and can directly predict flow field characteristics to give velocity cloud maps, and because there is no need to predict flow field information point by point, it is more concise and efficient, with lower information cost and higher timeliness. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a flow chart of an embodiment of the present invention;

[0068] Figure 2 This is a schematic diagram of the standard U-Net neural network architecture;

[0069] Figure 3 To improve the U-Net neural network architecture diagram;

[0070] Figure 4 This is a schematic diagram of the U-Net-LSTM neural network architecture;

[0071] Figure 5 The geometric image of the airfoil in the embodiment of the present invention;

[0072] Figure 6 for Figure 5 The corresponding SDF image;

[0073] Figure 7 The actual value of the steady flow field of the airfoil and its prediction result in the embodiment of the present invention;

[0074] Figure 8 It is the actual value of the unsteady flow field of the airfoil in the embodiment of the present invention and its predicted result. DETAILED DESCRIPTION

[0075] In order to make the technical problems, technical solutions and beneficial effects solved by the present invention clearer and more understandable, and to enable those skilled in the art to better understand the solutions of the present invention, the present invention is further described and fully described in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0076] The present invention is described below in conjunction with specific embodiments. Figure 1As shown, this embodiment uses a flow field image driven airfoil steady and unsteady aerodynamic modeling method to predict the steady and unsteady flow fields around a two-dimensional airfoil, including the following steps:

[0077] Step 1: Select a sample airfoil group and construct a sample data set.

[0078] When selecting a sample airfoil group, it is necessary to judge based on the airfoil to be predicted. If the type of the airfoil to be predicted is known, the sample airfoils in the sample airfoil group are consistent with the type of the airfoil to be predicted; if the airfoil to be predicted is not an airfoil of a known type, but an airfoil obtained by optimizing a basic airfoil of a known type, the sample airfoils in the sample airfoil group are consistent with the type of the basic airfoil. In this field, the type of airfoil is a concept well known to those skilled in the art. Typical airfoil types include NACA four-digit airfoils, NACA five-digit airfoils, RAE airfoils, etc. In this embodiment, 33 NACA series four-digit airfoils are selected as the sample airfoil group.

[0079] For a sample airfoil in the sample airfoil group, calculate the following data:

[0080] (1) Based on the CFD method, velocity cloud diagrams of the sample airfoil under several steady states and a sequence of velocity cloud diagrams under unsteady states are calculated; the steady state consists of a set angle of attack and a set incoming flow Mach number; the unsteady state consists of a set angle of attack variation law and a set incoming flow Mach number.

[0081] The specific process is:

[0082] For the steady state, in the given angle of attack range [ , ] and the incoming flow Mach number range [ , ], randomly select M angles of attack and L incoming Mach numbers Based on the CFD method, the velocity cloud diagram of the sample airfoil under the conditions of set angle of attack and set incoming flow Mach number is calculated, and a total of Velocity contours; thus creating a velocity contour for a sample airfoil training cases, created for each of the N airfoils in the sample airfoil group training cases;

[0083] For the unsteady state, the pitch oscillation of the sample airfoil around the 1 / 4 chord length point is taken as the research object, and the change process of its angle of attack is:

[0084]

[0085] Where: is the average angle of attack, is the angle of attack oscillation amplitude, is the oscillation frequency, the reduction frequency , where C is the reference length, V is the set flight speed, and the set reduction frequency range [ , ], select P reduction frequencies at equal intervals within the set reduction frequency range Based on the CFD method, the velocity cloud diagram sequence of the sample airfoil at the corresponding reduction frequency is calculated, and a total of P velocity cloud diagram sequences are obtained; a total of N airfoils in the sample airfoil group are created A sequence of training cases.

[0086] In this embodiment, for the steady state, the angle of attack range is set to [ , ], the incoming flow Mach number range is [ , ], for each sample airfoil, according to the given angle of attack and incoming flow Mach number range, 5 angles of attack and 2 incoming flow Mach numbers were randomly selected, and the velocity cloud diagram of the sample airfoil under the conditions of the set angle of attack and the set incoming flow Mach number was calculated based on the CFD method, and a velocity cloud diagram was created for each sample airfoil. training cases, a total of 33 sample airfoils were created training cases; in this embodiment, the selection rules and calculation conditions of the sample airfoil are shown in Table 1;

[0087] Table 1 Airfoil selection rules and calculation conditions

[0088]

[0089] For the unsteady state, for the above 33 sample airfoils, the pitch oscillation of the sample airfoil around the 1 / 4 chord length point is taken as the research object, and the change process of its angle of attack is:

[0090]

[0091] Where: is the average angle of attack, is the angle of attack oscillation amplitude, is the oscillation frequency, the reduction frequency , where C is the reference length, V is the set flight speed, and the reduced frequency range is determined [ , ], for each sample airfoil, 16 reduction frequencies were selected in the reduction frequency range with equal intervals of 0.002. Based on the CFD method, the velocity cloud diagram sequence of the sample airfoil at the corresponding reduction frequency was calculated. A total of 528 training cases were created for the 33 sample airfoils.

[0092] (2) According to the geometric image of the sample airfoil, the signed distance field image of the sample airfoil is obtained.

[0093] The specific process is:

[0094] Create a zero level set to represent the point set on the airfoil surface, denoted as Z, as follows:

[0095]

[0096] in, Indicates the plane area where the sample airfoil geometry image is located The point on is a level set function if and only if When located on the airfoil geometry boundary , if and only if Inside the airfoil geometry , if and only if Outside the airfoil geometry ;

[0097] Then, based on the level set function, the signed distance function is calculated as follows:

[0098]

[0099] Point Set The point on for The symbol; according to The value of is mapped to obtain the SDF color image. In this embodiment, Figure 5 and Figure 6 Shown are the airfoil geometry and the calculated SDF images.

[0100] (3) based on the geometric image of the sample airfoil and a certain constant state, calculating a synthetic image of the angle of attack-Mach number corresponding to the certain constant state;

[0101] Based on the geometric image of the sample airfoil, a certain angle of attack and a set incoming flow Mach number condition under an unsteady state, a synthetic image of the angle of attack-Mach number corresponding to the angle of attack and the set incoming flow Mach number condition is calculated.

[0102] The specific process of calculating the composite image of angle of attack-Mach number is as follows:

[0103] The pixel value in the synthetic image is recorded as , the calculation expression is as follows:

[0104]

[0105] in is the pixel value of the corresponding pixel point in the airfoil geometry image, ranging from [0,255], is the maximum value of pixels in the airfoil geometry image; is the incoming flow Mach number corresponding to the synthetic image, which is given by the calculation conditions, that is, the incoming flow Mach number in a certain steady state, or the set incoming flow Mach number in an unsteady state. is the maximum Mach number set.

[0106] Through the above calculations, the sample data corresponding to the sample airfoil are obtained as follows: the signed distance field image of the sample airfoil, the velocity cloud diagrams of the sample airfoil under several steady states and the synthetic images of the angle of attack-Mach number corresponding to the steady state, the velocity cloud diagram sequence of the sample airfoil under the unsteady state and the synthetic images of the angle of attack-Mach number corresponding to each angle of attack and the set incoming flow Mach number conditions under the unsteady state.

[0107] Step 2: Construct a U-Net neural network for steady flow field images; the U-Net neural network is an improved neural network based on the standard U-Net network model, with an input layer, a first encoder, a first decoder and an output layer as component modules; the standard U-Net network model architecture is as follows Figure 2 As shown, the improved U-Net neural network architecture of the present invention is as follows Figure 3 shown.

[0108] The input data of the U-Net neural network is the signed distance field image of the airfoil and the composite image of the angle of attack-Mach number under the steady state, and the output data is the velocity cloud map of the airfoil under the steady state.

[0109] The U-Net neural network processes input data as follows:

[0110] First, the input layer receives the input data and then passes it to the first encoder;

[0111] The first encoder includes multiple convolutional layers and maximum pooling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, a maximum pooling layer is connected to perform a maximum pooling operation on the input feature map, which is then used as the input data of the next convolutional layer; the feature map obtained by the last convolutional layer of the first encoder is input to the first decoder;

[0112] The first decoder includes multiple convolutional layers and upsampling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as input data for the next convolutional layer; after multiple convolutional layers, an upsampling layer is connected to perform a bilinear interpolation upsampling operation on the input feature map, and the feature map obtained after the upsampling operation is spliced ​​with the feature map before the corresponding maximum pooling operation in the first encoder, that is, the high-resolution feature map before the maximum pooling operation in the first encoder is spliced ​​with the feature map obtained after the upsampling operation using a skip connection operation as input data for the next convolutional layer, so that the spatial dimension of the feature map is restored layer by layer until the feature map obtained by the last convolutional layer is output to the output layer;

[0113] The output layer performs a 1×1 convolution operation on the input high-resolution feature map to obtain the final output data.

[0114] In this embodiment, the convolution operation performed by the convolution layer on the input data is a convolution operation with a padding operation. The padding operation ensures that the height and width of the feature map remain unchanged after each convolution, so that when the feature map before the maximum pooling operation is spliced ​​with the feature map obtained after the upsampling operation, no additional cropping operation is required; the padding operation is specifically as follows:

[0115] Assuming the size of the input feature map of the convolution layer is H×W, the size of the convolution kernel is K×K, the number of padding is P, and the stride is S, the size of the output feature map is:

[0116] +1

[0117] +1

[0118] The value of the padding number P is calculated based on the requirement that the size of the output feature map remains the same as the size of the input feature map, thereby ensuring that the flow field image feature map information is not lost.

[0119] In this embodiment, after the convolution layer performs a convolution operation on the input data, a batch normalization method is used to normalize the distribution of each feature channel in the feature map obtained by the convolution operation; the specific calculation formula of batch normalization is:

[0120]

[0121] in, Express Perform batch normalization operation. is the input data, is the batch mean, is the standard deviation of the batch, is the scaling parameter, is the translation parameter.

[0122] In this embodiment, the specific interpolation formula for the upsampling layer to perform a bilinear interpolation upsampling operation on the input data is as follows:

[0123]

[0124] in, Representing coordinates The pixel value after upsampling is is the coordinate in the feature map before upsampling The pixel values ​​of the four nearest neighboring pixels, that is, The pixel values ​​in the high-resolution feature map, and is the pixel value in the low-resolution feature map.

[0125] Step 3: Construct a U-Net-LSTM neural network for unsteady flow field images; the U-Net-LSTM neural network adds an LSTM module on the basis of the U-Net neural network established in step 2, and is composed of an input layer, a second encoder, a second decoder, an LSTM module and an output layer; specifically, the LSTM module is connected between the convolution layer before the maximum pooling operation in the second encoder and the convolution layer after the upsampling operation in the second decoder, and the multi-scale time series features are transmitted through the LSTM module.

[0126] The input data of the U-Net-LSTM neural network is a signed distance field image of the airfoil and an image sequence composed of multiple synthetic images of angle of attack-Mach number, and the output data is a velocity cloud map of the airfoil at the next moment in an unsteady state; the image sequence is a sequence of equal time intervals.

[0127] The U-Net-LSTM neural network processes the input data as follows:

[0128] First, the input layer receives the input data and then passes it to the second encoder;

[0129] The second encoder includes multiple convolutional layers and maximum pooling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, a maximum pooling layer is connected to perform a maximum pooling operation on the input feature map, which is then used as the input data of the next convolutional layer; the feature map obtained by the last convolutional layer of the second encoder is input to the second decoder;

[0130] The feature map output by the convolution layer before the maximum pooling operation in the second encoder is also input into the LSTM module, through which the feature map is processed in the time dimension to extract the feature map in the time series;

[0131] The second decoder includes multiple convolutional layers and upsampling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, an upsampling layer is connected to perform a bilinear interpolation upsampling operation on the input feature map, and the feature map obtained by the bilinear interpolation upsampling operation is concatenated with the feature map output by the corresponding LSTM module as the input data of the next convolutional layer, so that the spatial dimension of the feature map is restored layer by layer until the feature map obtained by the last convolutional layer is output to the output layer;

[0132] The output layer performs a 1×1 convolution operation on the input high-resolution feature map to obtain the final output data.

[0133] Step 4: Based on the sample data set obtained in step 1, a steady sample set is determined, and the U-Net neural network obtained in step 2 is trained using the steady sample set; based on the sample data set obtained in step 1, an unsteady sample set is determined, and the U-Net-LSTM neural network obtained in step 3 is trained using the unsteady sample set;

[0134] The steady sample set refers to a sample data set consisting of a signed distance field image of a sample airfoil, velocity cloud images of the sample airfoil under several steady states, and a synthetic image of angle of attack-Mach number corresponding to the steady state;

[0135] The unsteady sample set refers to a sample data set consisting of a signed distance field image of a sample airfoil, a velocity cloud map sequence of the sample airfoil in an unsteady state, and a synthetic image of the angle of attack-Mach number corresponding to various angles of attack and set incoming flow Mach number conditions in an unsteady state;

[0136] In this embodiment, the 330 steady-state training cases obtained in step 1 are divided into a training data set, a verification data set, and a test data set in a ratio of 8:1:1, and the U-Net neural network obtained in step 2 is trained, verified, and tested. The 528 groups of unsteady-state training cases obtained in step 1 are also divided into a training data set, a verification data set, and a test data set in a ratio of 8:1:1, and the U-Net-LSTM neural network obtained in step 3 is trained, verified, and tested.

[0137] Step 5: Rapidly predict the steady and unsteady flow fields of the airfoil to be predicted;

[0138] For steady flow prediction:

[0139] Obtain a signed distance field image of the airfoil to be predicted, input the signed distance field image of the airfoil to be predicted and a synthetic image of the angle of attack-Mach number under a set steady state into a trained U-Net neural network, obtain a velocity cloud map of the airfoil to be predicted under a set steady state, and realize steady flow field prediction of the airfoil to be predicted;

[0140] For unsteady flow prediction:

[0141] The signed distance field image of the airfoil to be predicted is obtained, and an image sequence consisting of the signed distance field image of the airfoil to be predicted and a synthetic image of the angle of attack-Mach number under a set unsteady state is input into a trained U-Net-LSTM neural network to obtain the velocity cloud map of the airfoil to be predicted at the next moment under the set unsteady state, thereby realizing the unsteady flow field prediction of the airfoil to be predicted.

[0142] Figure 7 The steady-state prediction results of the U-Net neural network are shown. Figure 7 In the figure, from top to bottom, the predicted results and the actual values ​​of the airfoil shown in Table 2 are compared under the shown conditions.

[0143] Table 2 Verified airfoil and predicted state

[0144]

[0145] like Figure 7 As shown, the left column is the true value, the middle column is the predicted value, and the right column is the error value between the two. The error calculation formula is as follows:

[0146]

[0147] in is the nth real image, is the nth predicted image, N is the total number of pixels in the image, is the set of image pixels, are pixel coordinates;

[0148] Depend on Figure 7 It can be seen that the maximum error of the steady prediction example is 1.06. In the RGB image prediction problem, when the mean square error value is less than 5, the image difference is small, which belongs to high-precision prediction. It can be seen that the neural network model can accurately capture the characteristics of the steady flow field and predict the steady flow field around the airfoil at low Reynolds number with high accuracy.

[0149] Figure 8 The prediction results of the U-Net-LSTM neural network are shown. Figure 8The figure shows the comparison between the predicted results and the actual values ​​of the velocity field at 0.1s, 0.2s, 0.3s, and 0.4s for NACA0015 airfoil and NACA2415 airfoil when the reduction frequency K=0.08. Figure 8 The top two rows correspond to the NACA0015 airfoil, and the bottom two rows correspond to the NACA2415 airfoil. Judging from the prediction results at different times, the constructed U-net-LSTM image modeling method also has good adaptability to unsteady time series image data, with the maximum prediction error of 1.96. It can be seen that the U-Net-LSTM neural network can accurately capture the time series variation characteristics of the unsteady flow field and predict the airfoil pitch oscillation flow field with high accuracy. In the aerodynamic design process, different design schemes can be quickly iterated, reducing the need for a large number of CFD calculations, thereby greatly improving design efficiency.

[0150] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention and without making creative work. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A flow field image driven airfoil steady and unsteady aerodynamic modeling method, characterized by: The steps include: Step 1: Select a sample airfoil group and construct a sample data set; For a sample airfoil in the sample airfoil group: Based on the CFD method, velocity cloud diagrams of the sample airfoil under several steady states and a sequence of velocity cloud diagrams under unsteady states are calculated; the steady state is composed of a set angle of attack and a set incoming flow Mach number; the unsteady state is composed of a set angle of attack variation law and a set incoming flow Mach number; According to the geometric image of the sample airfoil, a signed distance field image of the sample airfoil is obtained; According to the geometric image of the sample airfoil and a certain constant state, a synthetic image of the angle of attack-Mach number corresponding to the certain constant state is calculated; According to the geometric image of the sample airfoil, a certain angle of attack and a set incoming flow Mach number condition under an unsteady state, a synthetic image of the angle of attack-Mach number corresponding to the angle of attack and the set incoming flow Mach number condition is calculated; The sample data corresponding to the sample airfoil are: a signed distance field image of the sample airfoil, velocity cloud images of the sample airfoil under several steady states and a composite image of the angle of attack-Mach number corresponding to the steady state, a sequence of velocity cloud images of the sample airfoil under an unsteady state and a composite image of the angle of attack-Mach number corresponding to various angles of attack and set incoming flow Mach number conditions under the unsteady state; Step 2: constructing a U-Net neural network for steady flow field images; the U-Net neural network is composed of an input layer, a first encoder, a first decoder and an output layer; the input data of the U-Net neural network is a signed distance field image of the airfoil and a composite image of the angle of attack-Mach number, and the output data is a velocity cloud map of the airfoil in a steady state; Step 3: construct a U-Net-LSTM neural network for unsteady flow field images; the U-Net-LSTM neural network is a U-Net neural network established in step 2 with an LSTM module added, and is composed of an input layer, a second encoder, a second decoder, an LSTM module, and an output layer; the input data of the U-Net-LSTM neural network is a signed distance field image of the airfoil, and an image sequence composed of a plurality of synthetic images of angle of attack-Mach number, and the output data is a velocity cloud map of the airfoil at the next moment in an unsteady state; the image sequence is a sequence of equal time intervals; Step 4: Based on the sample data set obtained in step 1, a steady sample set is determined, and the U-Net neural network obtained in step 2 is trained using the steady sample set; based on the sample data set obtained in step 1, an unsteady sample set is determined, and the U-Net-LSTM neural network obtained in step 3 is trained using the unsteady sample set; The steady sample set refers to a sample data set consisting of a signed distance field image of a sample airfoil, velocity cloud images of the sample airfoil under several steady states, and a synthetic image of angle of attack-Mach number corresponding to the steady state; The unsteady sample set refers to a sample data set consisting of a signed distance field image of a sample airfoil, a velocity cloud map sequence of the sample airfoil in an unsteady state, and a synthetic image of the angle of attack-Mach number corresponding to various angles of attack and set incoming flow Mach number conditions in an unsteady state; Step 5: Rapidly predict the steady and unsteady flow fields of the airfoil to be predicted; For steady flow prediction: Obtain a signed distance field image of the airfoil to be predicted, input the signed distance field image of the airfoil to be predicted and a synthetic image of the angle of attack-Mach number under a set steady state into a trained U-Net neural network, obtain a velocity cloud map of the airfoil to be predicted under a set steady state, and realize steady flow field prediction of the airfoil to be predicted; For unsteady flow prediction: The signed distance field image of the airfoil to be predicted is obtained, and an image sequence consisting of the signed distance field image of the airfoil to be predicted and a synthetic image of the angle of attack-Mach number under a set unsteady state is input into a trained U-Net-LSTM neural network to obtain the velocity cloud map of the airfoil to be predicted at the next moment under the set unsteady state, thereby realizing the unsteady flow field prediction of the airfoil to be predicted.

2. The method for modeling steady and unsteady aerodynamic forces of an airfoil driven by flow field images according to claim 1, characterized in that: If the type of the airfoil to be predicted is known, the sample airfoils in the sample airfoil group are consistent with the type of the airfoil to be predicted; if the airfoil to be predicted is an airfoil obtained by optimizing the basic airfoil, the sample airfoils in the sample airfoil group are consistent with the type of the basic airfoil.

3. The method for modeling steady and unsteady aerodynamic forces of an airfoil driven by flow field images according to claim 1, characterized in that: In step 1, the process of calculating the velocity cloud diagrams of the sample airfoil under several steady states and the sequence of velocity cloud diagrams under unsteady states based on the CFD method is as follows: For the steady state, in the given angle of attack range [ , ] and the incoming flow Mach number range [ , ], randomly select M angles of attack and L incoming Mach numbers , based on the CFD method, the velocity cloud diagram of the sample airfoil under the conditions of set angle of attack and set incoming flow Mach number is calculated; For the unsteady state, the pitch oscillation of the sample airfoil around the 1 / 4 chord length point is taken as the research object, and the change process of its angle of attack is: Where: is the average angle of attack, is the angle of attack oscillation amplitude, is the oscillation frequency, is the time; reduce the frequency , where C is the reference length, V is the set flight speed, and the set reduction frequency range [ , ], select P reduction frequencies at equal intervals within the set reduction frequency range , based on the CFD method, the velocity contour sequence of the sample airfoil at the corresponding reduced frequency is calculated.

4. The method for modeling steady and unsteady aerodynamic forces of an airfoil driven by flow field images according to claim 1, characterized in that: In step 1, the process of obtaining the signed distance field image of the sample airfoil according to the geometric image of the sample airfoil is as follows: Create a zero level set to represent the point set on the airfoil surface, denoted as Z, as follows: in, Indicates the plane area where the sample airfoil geometry image is located The point on is a level set function if and only if When located on the airfoil geometry boundary , if and only if When inside the airfoil geometry boundary , if and only if Outside the airfoil geometry boundary ; Then, based on the level set function, the signed distance function is calculated as follows: Point Set The point on for The symbol; according to The numerical value of is mapped to obtain the SDF color image.

5. The method for modeling steady and unsteady aerodynamic forces of an airfoil driven by flow field images according to claim 1, characterized in that: In step 1, the process of calculating the composite image of angle of attack-Mach number is: The pixel value in the synthetic image is recorded as , the calculation expression is as follows: in is the pixel value of the corresponding pixel point in the airfoil geometry image, is the maximum value of pixels in the airfoil geometry image; is the incoming flow Mach number corresponding to the synthetic image, is the maximum Mach number set.

6. The method for modeling steady and unsteady aerodynamic forces of an airfoil driven by flow field images according to claim 1, characterized in that: The U-Net neural network constructed in step 2 processes the input data as follows: First, the input layer receives input data and then passes it to the first encoder; The first encoder includes multiple convolutional layers and maximum pooling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, a maximum pooling layer is connected to perform a maximum pooling operation on the input feature map, which is then used as the input data of the next convolutional layer; the feature map obtained by the last convolutional layer of the first encoder is input to the first decoder; The first decoder includes multiple convolutional layers and upsampling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, an upsampling layer is connected to perform a bilinear interpolation upsampling operation on the input feature map, and the feature map obtained after the upsampling operation is concatenated with the feature map before the corresponding maximum pooling operation in the first encoder as the input data of the next convolutional layer, so that the spatial dimension of the feature map is restored layer by layer until the feature map obtained by the last convolutional layer is output to the output layer; The final output data is obtained through the output layer.

7. The method for modeling steady and unsteady aerodynamic forces of an airfoil driven by flow field images according to claim 1, characterized in that: The U-Net-LSTM neural network constructed in step 3 introduces the LSTM module on the basis of the U-Net neural network constructed in step 2. Specifically, the LSTM module is connected between the convolution layer before the maximum pooling operation in the second encoder and the convolution layer after the upsampling operation in the second decoder, and the multi-scale time series features are transmitted through the LSTM module.

8. The method for modeling steady and unsteady aerodynamic forces of an airfoil driven by flow field images according to claim 7, characterized in that: The U-Net-LSTM neural network constructed in step 3 processes the input data as follows: First, the input layer receives the input data and then passes it to the second encoder; The second encoder includes multiple convolutional layers and maximum pooling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, a maximum pooling layer is connected to perform a maximum pooling operation on the input feature map, which is then used as the input data of the next convolutional layer; the feature map obtained by the last convolutional layer of the second encoder is input to the second decoder; The feature map output by the convolution layer before the maximum pooling operation in the second encoder is also input into the LSTM module, through which the feature map is processed in the time dimension to extract the feature map in the time series; The second decoder includes multiple convolutional layers and upsampling layers; each convolutional layer performs a convolution operation on the input data to obtain a feature map of the input data as the input data of the next convolutional layer; after multiple convolutional layers, an upsampling layer is connected to perform a bilinear interpolation upsampling operation on the input feature map, and the feature map obtained by the bilinear interpolation upsampling operation is concatenated with the feature map output by the corresponding LSTM module as the input data of the next convolutional layer, so that the spatial dimension of the feature map is restored layer by layer until the feature map obtained by the last convolutional layer is output to the output layer; The final output data is obtained through the output layer.

9. The method for modeling steady and unsteady aerodynamic forces of an airfoil driven by flow field images according to claim 6, characterized in that: The convolution operation performed by the convolution layer on the input data is a convolution operation with a padding operation. The padding operation ensures that the height and width of the feature map remain unchanged after each convolution.

10. The method for modeling steady and unsteady aerodynamic forces of an airfoil driven by flow field images according to claim 6, characterized in that: After the convolution layer performs a convolution operation on the input data, the batch normalization method is used to normalize the distribution of each feature channel in the feature map obtained by the convolution operation.

Images