A flow field prediction method based on dense convolutional network

Abstract

The present invention discloses a flow field prediction method based on a dense convolutional network, comprising the following steps: obtaining an airfoil shape data set and a simulated flow field data set; obtaining airfoil geometric parameters using a dense convolutional network with an adaptive Dropout module; constructing an airfoil parameter and physical parameter input use case; training a flow field prediction model based on MHP; and predicting the flow field on different airfoil data using the trained flow field prediction model. The present invention adds a SeLU activation function and an adaptive Dropout when using a dense convolutional network to extract airfoil feature data, thereby alleviating the overfitting of the flow field prediction method due to the small amount of data, and enhancing the prediction accuracy and generalization of the neural network; at the same time, a multi-head perceptron is used to train the prediction model and perform prediction, thereby avoiding the interference of sparse data on other aerodynamic parameters to be predicted, making up for the defects of MLP when processing sparse data, and improving the prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of flow field prediction, and in particular relates to a flow field prediction method based on a dense convolutional network. Background Art

[0002] Airfoil is an important fluid mechanics software widely used in aerospace, energy and power and other fields. The flow field prediction of airfoil refers to the use of numerical simulation technology to predict and analyze the flow behavior of airfoil under different wind tunnel or airflow conditions. It is one of the important processes for evaluating aircraft design and airfoil performance. Airfoil flow field prediction has important applications in aerospace, aircraft design, wind power generation, automotive engineering and other fields. It can provide key information about airfoil performance to guide product design and performance improvement. The initial method of obtaining airfoil flow field data was to obtain results through wind tunnel tests, but this method requires a large amount of time and economic cost and is often used in the later stages of aircraft design. The computational fluid dynamics (CFD) method that emerged later can be used to obtain the flow field of airfoils. This method requires the use of solving software to iteratively calculate and perform high-precision numerical simulations, such as solving the Euler equations, Reynolds-averaged Navier-Stokes (RANS) equations, etc. However, this method also takes a long time and has high requirements for hardware.

[0003] With the development of deep learning, neural networks are used to construct regression prediction models to automatically learn high-dimensional latent mapping relationships and achieve rapid prediction of flow fields. This can approximate the numerical simulation results of the solution software to a certain extent, thus providing convenience for researchers. However, the existing deep learning neural networks cannot effectively extract the complex features of the airfoil, which makes the flow field prediction accuracy low. At the same time, a large amount of data is required to train the neural network prediction model. The high requirements for data have become the main obstacle to the practical application of these models. When the data is insufficient, the model usually has performance problems such as overfitting. Therefore, how to design a flow field prediction method that can effectively extract airfoil features and improve the accuracy and generalization of airfoil flow field prediction on a small-scale data set is an urgent problem to be solved. Summary of the invention

[0004] The technical problem to be solved by the present invention is a flow field prediction method based on a dense convolutional network, which improves the generalization and accuracy of flow field prediction training on small-scale data sets, and alleviates the problem of overfitting in the flow field prediction field due to small data amounts.

[0005] In order to solve the above technical problems, the present invention is implemented in the following ways:

[0006] A flow field prediction method based on a dense convolutional network comprises the following steps:

[0007] S1, obtaining an airfoil shape data set and a simulated flow field data set, and preprocessing the airfoil shape data;

[0008] S2, using a dense convolutional network with an adaptive Dropout module to obtain airfoil geometric parameters;

[0009] S3, construct a use case for inputting airfoil parameters and physical parameters;

[0010] S4, training the flow field prediction model based on MHP;

[0011] S5. Use the trained flow field prediction model to predict the flow field on different airfoil data.

[0012] Furthermore, the specific steps of step S1 are as follows:

[0013] S11, obtaining a reference airfoil from an airfoil shape data set and fitting it, and obtaining coordinate details on the fitting curve;

[0014] S12, the acquired coordinates generate an inverted grayscale image of the airfoil, and set the pixels that completely pass through the airfoil contour to 1, the pixels that do not pass through the airfoil contour to 0, and the remaining pixel values ​​are set between 0 and 1.

[0015] Furthermore, the specific steps of step S2 are as follows:

[0016] S21, for the generated airfoil inverted grayscale image X∈R C*H*W , and input it into the convolutional layer for image dimensionality reduction, where R represents the real number domain, C represents the number of channels of the input image, and H and W represent the height and width of the feature map; the output image Among them C out Indicates the number of output channels, H out and W out Represents the height and width of the output feature map;

[0017] S22. The output of the convolution layer is represented as F, which is alternately passed through three DenseBlocks and two Transition Layers of the dense convolutional network, and then pooled and linearly operated to obtain the airfoil geometric parameters.

[0018] Furthermore, the specific steps of step S22 are as follows:

[0019] S221, the output F of the convolutional layer enters the first layer Denselayer in DenseBlock;

[0020] The Batch Normalization Layer in S222 and Denselayer performs batch normalization to obtain the normalized feature map Y bn ;

[0021] S223, through the activation function SeLU feature map Y bn Perform element-by-element correction to obtain the corrected feature map Y SeLU , the SeLU function calculation formula is as follows:

[0022]

[0023] S224, corrected feature map Y SeLU Perform a 1×1 convolution operation to obtain the convolved feature map Y conv ;

[0024] S225, feature map Y conv After the Batch Normalization Layer and the activation function SeLU again, we get Y SeLU′ , and perform a 3×3 convolution operation on it to obtain the feature map Y conv′ ;

[0025] S226, for the feature map Y obtained after convolution conv′ Perform adaptive Drop and normalize it to the maximum and minimum to get Y (n) ; Use Bernoulli distribution Bernoulli (γ) for random sampling to obtain a number of pixel points M i，j , for each pixel M i，j , a pixel block of size block_size×block_size is set as the center. In each pixel block, the first k percentile elements are set to 0 and the remaining elements are set to 1, thereby generating a template M, Y′ (n) The expression is as follows:

[0026] Y′ (n) =Y (n) ×M

[0027] Then for Y′ (n) For scaling output, the expression is as follows:

[0028] Y′ (n+1) =Y′ (n) ×count(M)÷count_ones(M)

[0029] Among them, count(M) represents the number of elements in template M, count_ones(M) represents the number of 1s in template M, and γ represents the number of features that need to be dropped. The expression is as follows:

[0030]

[0031] Among them, keep_prob represents the probability of retaining the unit block; feat_size represents the feature map size;

[0032] S227, the output Y′ of the first layer Denselayer (n+1) It is input into the subsequent three layers of Denselayer in sequence, and the input of each layer of Denselayer is obtained by combining the features output by all the previous Denselayer layers. The expression is as follows:

[0033] x ι =H ι ([x0, x1, x2...x ι-1 ])

[0034] Among them, x0, x1, x2...x l-1 H refers to the concatenation of feature maps generated in layers 0, 1, ..., ι-1. ι (·) function represents the process of extracting features by the ι-layer Denselayer;

[0035] S228, the output of the fourth layer Denselayer, that is, the feature map obtained by the first DenseBlock is input to the Transition Layer, and the feature channel dimension is reduced by a 1×1 convolution kernel, and then the feature is downsampled by a 2×2 average pooling layer; and then it continues to be input to the next DenseBlock, Transition Layer and DenseBlock;

[0036] S229. The feature map Y output from the last Denseblock module is passed through the pooling layer and the linear layer to obtain the airfoil feature parameter P.

[0037] Furthermore, the specific method of step S3 is as follows:

[0038] The characteristic parameter P obtained by the dense convolutional network is combined with the coordinates (x, y) of the airfoil control point, the Reynolds number (Re) and the angle of attack (AOA) to form an input case, and a data set for training the flow field prediction model is constructed with the flow field data of CFD simulation.

[0039] Furthermore, the specific steps of step S4 are as follows:

[0040] S41, using the airfoil characteristic parameters obtained in step S2 and the physical characteristics constructed in step S3 as input, training the model, and using a multi-head network to obtain the model's prediction output for physical parameters (u, p, v) with different distribution characteristics, the specific expression of which is as follows:

[0041]

[0042] Among them, the left side of the formula represents the prediction model of MHP, u, p, and v represent the x-direction velocity component, pressure value, and y-direction velocity component of the flow field respectively;

[0043] S42. The LOSS function expression of each head in the multi-head perceptron is as follows:

[0044]

[0045] in, They represent the x-direction velocity component, pressure value and y-direction velocity component of the flow field simulated by CFD, respectively. Respectively represent the predicted values;

[0046] S43. When LOSS converges, the training ends and the prediction model is obtained.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] The present invention extracts airfoil features by adding a dense convolutional network with an adaptive Dropout module, which alleviates the overfitting of the flow field prediction method caused by the small amount of data, helps to improve the regularization ability of the network, and enhances the prediction accuracy and generalization of the neural network; at the same time, the multi-head perceptron is used to train the prediction model and make predictions, avoiding the interference of sparse data on other aerodynamic parameters to be predicted, making up for the defects of MLP in processing sparse data, and improving the prediction accuracy. By setting the Drop probability, important feature information can be retained, avoiding performance loss caused by excessive feature discarding, and the series feature allows the feature to be fully reused. The high-level and low-level feature fusion enhances the network's anti-overfitting performance, reduces parameters, and alleviates the gradient disappearance problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the flow chart of the flow field prediction method of the present invention. DETAILED DESCRIPTION

[0050] The specific implementation of the present invention is further described in detail below with reference to the accompanying drawings and specific examples.

[0051] like Figure 1 As shown, a flow field prediction method based on a dense convolutional network includes the following steps:

[0052] S1. Obtain an airfoil shape data set and a simulated flow field data set, and preprocess the airfoil shape data. The specific steps are as follows:

[0053] S11, obtaining a reference airfoil from an airfoil shape data set and fitting it, and obtaining coordinate details on the fitting curve;

[0054] S12, the acquired coordinates generate an inverted grayscale image of the airfoil, and set the pixels that completely pass through the airfoil contour to 1, the pixels that do not pass through the airfoil contour to 0, and the remaining pixel values ​​are set between 0 and 1.

[0055] S2. Use a dense convolutional network with an adaptive Dropout module to obtain airfoil geometric parameters. The specific steps are as follows:

[0056] S21, for the generated airfoil inverted grayscale image X∈R C*H*W , and input it into the convolutional layer for image dimensionality reduction, where R represents the real number domain, C represents the number of channels of the input image, and H and W represent the height and width of the feature map; the output image Among them C out Indicates the number of output channels, H out and W out Represents the height and width of the output feature map;

[0057] S22. The output of the convolution layer is represented as F, which is alternately passed through three DenseBlocks and two Transition Layers (DenseBlock, Transition Layer, ..., DenseBlock) of the dense convolutional network, and then pooled and linearly operated to obtain the airfoil geometric parameters.

[0058] The specific steps of step S22 are as follows:

[0059] S221, the output F of the convolutional layer enters the first layer Denselayer in DenseBlock;

[0060] The Batch Normalization Layer in S222 and Denselayer performs batch normalization to obtain the normalized feature map Y bn ;

[0061] S223, through the activation function SeLU (scaled exponential linear unit) on the feature map Y bn Perform element-by-element correction to obtain the corrected feature map Y SeLU , the SeLU function calculation formula is as follows:

[0062]

[0063] S224, corrected feature map Y SeLU Perform a 1×1 convolution operation to obtain the convolved feature map Y conv ;

[0064] S225, feature map Y conv After the Batch Normalization Layer and the activation function SeLU again, we get Y SeLU′ , and perform a 3×3 convolution operation on it to obtain the feature map Y conv′ ;

[0065] S226, for the feature map Y obtained after convolution conv Perform adaptive Drop and normalize it to the maximum and minimum to get Y (n) ; Use Bernoulli distribution Bernoulli (γ) for random sampling to obtain a number of pixel points M i，j , for each pixel M i，j , a pixel block of size block_size×block_size is set as the center. In each pixel block, the first k percentile elements are set to 0 and the remaining elements are set to 1, thereby generating a template M, Y′ (n) The expression is as follows:

[0066] Y′ (n) =Y (n) ×M

[0067] Then for Y′ (n) For scaling output, the expression is as follows:

[0068] Y′ (n+1) =Y′ (n) ×count(M)÷count_ones(M)

[0069] Among them, count(M) represents the number of elements in template M, count_ones(M) represents the number of 1s in template M, block_size is generally 3, 5, 7. If it is 1, the algorithm will degenerate into the traditional dropout module. k is generally between 30 and 50. γ represents the number of features that need to be dropped, and its expression is as follows:

[0070]

[0071] Among them, keep_prob represents the probability of retaining the unit block; feat_size represents the feature map size;

[0072] S227, the output Y′ of the first layer Denselayer(n+1) It is input into the subsequent three layers of Denselayer in sequence, and the input of each layer of Denselayer is obtained by combining the features output by all the previous Denselayer layers. The expression is as follows:

[0073] x ι =H ι ([x0, x1, x2...x ι-1 ])

[0074] Among them, x0,x1,x2…x ι-1 H refers to the concatenation of feature maps generated in layers 0, 1, ..., ι-1. ι (·) function represents the process of extracting features by the ι-layer Denselayer;

[0075] S228, the output of the fourth layer Denselayer, that is, the feature map obtained by the first DenseBlock is input to the Transition Layer, and the feature channel dimension is reduced by a 1×1 convolution kernel, and then the feature is downsampled by a 2×2 average pooling layer; and then it continues to be input to the next DenseBlock, Transition Layer and DenseBlock;

[0076] S229. The feature map Y output from the last Denseblock module is passed through the pooling layer and the linear layer to obtain the airfoil feature parameter P.

[0077] S3. Construct a use case for inputting airfoil parameters and physical parameters. The specific method is as follows:

[0078] The characteristic parameter P obtained by the dense convolutional network is combined with the coordinates (x, y) of the airfoil control point, the Reynolds number (Re), and the angle of attack (AOA) to form an input case, and a data set for training the flow field prediction model is constructed with the flow field data of CFD simulation.

[0079] S4. Train the flow field prediction model based on MHP. The specific steps are as follows:

[0080] S41, using the airfoil characteristic parameters obtained in step S2 and the physical characteristics constructed in step S3 as input, training the model, and using a multi-head network to obtain the model's prediction output for physical parameters (u, p, v) with different distribution characteristics, the specific expression of which is as follows:

[0081]

[0082] Among them, the left side of the formula represents the prediction model of MHP, u, p, and v represent the x-direction velocity component, pressure value, and y-direction velocity component of the flow field respectively;

[0083] S42. The LOSS function expression of each head in the multi-head perceptron is as follows:

[0084]

[0085] in, They represent the x-direction velocity component, pressure value and y-direction velocity component of the flow field simulated by CFD, respectively. Respectively represent the predicted values;

[0086] S43. When LOSS converges, the training ends and the prediction model is obtained.

[0087] S5. Use the trained flow field prediction model (MHP network) to input different airfoil data (airfoil characteristics, angle of attack, Reynolds number) to predict the flow field.

[0088] The above description is only an implementation mode of the present invention. It is stated again that for ordinary technicians in this technical field, several improvements can be made to the present invention without departing from the principle of the present invention. These improvements are also included in the protection scope of the claims of the present invention.

Claims

1. A flow field prediction method based on a dense convolutional network, characterized in that: The following steps are involved: S1, obtaining an airfoil shape data set and a simulated flow field data set, and preprocessing the airfoil shape data; S2, using a dense convolutional network with an adaptive Dropout module to obtain airfoil geometric parameters; S3, construct a use case for airfoil parameters and physical parameters input; S4, training the flow field prediction model based on MHP; S5. Predicting the flow field on different airfoil data using the trained flow field prediction model; The specific steps of step S1 are as follows: S11, obtaining a reference airfoil from an airfoil shape data set and fitting it, and obtaining coordinate details on the fitting curve; S12, the acquired coordinates generate an inverted grayscale image of the airfoil, and set the pixels that completely pass through the airfoil profile to 1, the pixels that do not pass through the airfoil profile to 0, and the remaining pixel values ​​are set between 0 and 1; The specific steps of step S2 are as follows: S21, for the generated airfoil inverted grayscale image X∈R C*H*W , and input it into the convolutional layer for image dimensionality reduction, where R represents the real number domain, C represents the number of channels of the input image, and H and W represent the height and width of the feature map; the output image Among them C out Indicates the number of output channels, H out and W out Represents the height and width of the output feature map; S22, the output of the convolution layer is represented as F, which is alternately passed through three DenseBlocks and two Transition Layers of the dense convolution network, and then pooled and linearly operated to obtain the airfoil geometric parameters; The specific steps of step S4 are as follows: S41, using the airfoil characteristic parameters obtained in step S2 and the physical characteristics constructed in step S3 as input, training the model, and using a multi-head network to obtain the model's prediction output for physical parameters (u, p, v) with different distribution characteristics, the specific expression of which is as follows: Among them, the left side of the formula represents the prediction model of MHP, u, p, and v represent the x-direction velocity component, pressure value, and y-direction velocity component of the flow field respectively; S42. The LOSS function expression of each head in the multi-head perceptron is as follows: in, They represent the x-direction velocity component, pressure value and y-direction velocity component of the flow field simulated by CFD, respectively. Respectively represent the predicted values; S43. When LOSS converges, the training ends and the prediction model is obtained.

2. The flow field prediction method based on a dense convolutional network according to claim 1, characterized in that: The specific method of step S3 is as follows: The characteristic parameter P obtained by the dense convolutional network is combined with the coordinates (x, y) of the airfoil control point, the Reynolds number (Re), and the angle of attack (AOA) to form an input case, and a data set for training the flow field prediction model is constructed with the flow field data of CFD simulation.

3. The flow field prediction method based on dense convolutional network according to claim 1, characterized in that: The specific steps of step S22 are as follows: S221, the output F of the convolutional layer enters the first layer Denselayer in DenseBlock; The Batch Normalization Layer in S222 and Denselayer performs batch normalization to obtain the normalized feature map Y bn ; S223, through the activation function SeLU feature map Y bn Perform element-by-element correction to obtain the corrected feature map Y SeLU , the SeLU function calculation formula is as follows: S224, corrected feature map Y SeLU Perform a 1×1 convolution operation to obtain the convolved feature map Y conv ; S225, feature map Y conv After the Batch Normalization Layer and the activation function SeLU again, we get Y SeLU' , and perform a 3×3 convolution operation on it to obtain the feature map Y conv' ; S226, for the feature map Y obtained after convolution conv’ Perform adaptive Drop and normalize it to the maximum and minimum to get Y (n) ; Use Bernoulli distribution Bernoulli (γ) for random sampling to obtain a number of pixel points M i,j , for each pixel M i,j , set a pixel block of size block_size×block_size with it as the center, in each pixel block, the first k percentile elements are set to 0, and the remaining elements are set to 1, thereby generating a template M, Y' (n) The expression is as follows: AND' (n) =And (n) ×M Then to Y' (n) For scaling output, the expression is as follows: AND' (n+1) =And' (n) ×count(M)÷count_ones(M) Among them, count(M) represents the number of elements in template M, count_ones(M) represents the number of 1s in template M, and γ represents the number of features that need to be dropped. The expression is as follows: Among them, keep_prob represents the probability of retaining the unit block; feat_size represents the feature map size; S227, the output Y' of the first layer Denselayer (n+1) It is input into the subsequent three layers of Denselayer in sequence, and the input of each layer of Denselayer is obtained by combining the features output by all the previous Denselayer layers. The expression is as follows: x ι =H ι ([x0,x1,x2…x ι-1 ]) Among them, x0,x1,x2…x ι-1 H refers to the concatenation of feature maps generated in layers 0, 1, ..., ι-1. ι (·) function represents the process of extracting features by the ι-layer Denselayer; S228, the output of the fourth layer Denselayer, the feature map obtained by the first DenseBlock is input to the TransitionLayer, the feature channel dimension is reduced by a 1×1 convolution kernel, and then the feature is downsampled by a 2×2 average pooling layer; and then it continues to be input to the next DenseBlock, Transition Layer and DenseBlock; S229. The feature map Y output from the last Denseblock module is passed through the pooling layer and the linear layer to obtain the airfoil feature parameter P.

Images