A method and system for intelligent reconstruction of unsteady flow field in wide-range internal rotation inlet

Abstract

The present invention discloses a method and system for intelligently reconstructing the unsteady flow field of a wide-area inward-rotating inlet, belonging to the field of supersonic flow technology. The method comprises the following steps: S1, performing CFD numerical simulation on a two-dimensional inward-rotating inlet at different Mach numbers to obtain data sets corresponding to different Mach numbers; S2, preprocessing the data sets corresponding to different Mach numbers; S3, outputting a predicted flow field; S4, generating a shock wave boundary layer separation point position detection data set; S5, outputting a predicted shock wave boundary layer separation point position; S6, visualizing the predicted flow field and the predicted shock wave boundary layer separation point position. The present invention achieves visualization of the flow field and shock wave boundary layer separation point position, which is of great significance for improving the starting capability of the inlet, promoting inlet design optimization, and ensuring stable engine operation.

Description

Technical Field

[0001] The present invention belongs to the field of supersonic flow technology, and in particular relates to a method and system for intelligently reconstructing an unsteady flow field of a wide-range internal rotating inlet. Background Art

[0002] Aerospace vehicles have garnered widespread attention worldwide due to their advantages, including short defense response times and strong penetration capabilities. As a key component of a scramjet engine, the inlet's primary function is to efficiently compress the captured airflow, providing the combustion chamber with air of a specific pressure, velocity, and uniformity, enabling the entire propulsion system to generate sufficient thrust to meet the Mach number requirements of a hypersonic vehicle. Among these, inward-turning inlets offer advantages such as high captured flow, high compression efficiency, a small wetted area, excellent off-design performance, and wide adaptability. However, the inherent geometry of inward-turning inlets presents complex and unique flow phenomena, such as three-dimensional curved surface shock wave / shock wave interference and curved surface shock wave / curved surface boundary layer interference, which significantly impact the inlet's starting performance. Traditional inlet flow field analysis relies primarily on discrete, low-density raw information collected by a single sensor, resulting in numerous limitations when dealing with wide-area unsteady flow fields.

[0003] In recent years, the rapid development of artificial intelligence (AI) technology has also sparked significant interest in the field of fluid dynamics. Unlike traditional artificial neural networks, spiking neural networks (SNs) possess rich spatiotemporal neurodynamic properties and diverse encoding mechanisms. In hardware, they offer advantages such as fast processing speed, low power consumption, ease of development, and high flexibility. By embedding SN weights into a hardware platform for brain-inspired computing, high-precision and efficient reconstruction of the unsteady flow field of a wide-area internal rotating inlet can be achieved, promoting the practical application of intelligent algorithms in the aviation field. Summary of the Invention

[0004] In order to solve the above problems, the present invention proposes a method and system for intelligent reconstruction of the unsteady flow field of a wide-range internal rotation inlet.

[0005] The technical solution of the present invention is: a method for intelligent reconstruction of unsteady flow field of wide-range internal rotation inlet includes the following steps:

[0006] S1. Perform CFD numerical simulation on the two-dimensional inward-turning inlet at different Mach numbers to obtain data sets corresponding to different Mach numbers.

[0007] S2. Preprocessing the data sets corresponding to different Mach numbers;

[0008] S3. Constructing an inward-turning inlet flow field reconstruction model, inputting the pre-processed data sets corresponding to different Mach numbers into the inward-turning inlet flow field reconstruction model, and outputting a predicted flow field;

[0009] S4. Obtain the distance from the shock wave boundary layer separation point to the two-dimensional inward-turning inlet inlet, and generate a shock wave boundary layer separation point position detection data set;

[0010] S5. Constructing a shock wave boundary layer separation point position detection model, inputting the shock wave boundary layer separation point position detection data set into the shock wave boundary layer separation point position detection model, and outputting a predicted shock wave boundary layer separation point position;

[0011] S6. Visualize the predicted flow field and the predicted shock wave boundary layer separation points.

[0012] Furthermore, in S1, the data sets corresponding to different Mach numbers include upper and lower wall pressure data, pressure field, density field, schlieren image, velocity field in the x-direction, and velocity field in the y-direction of the two-dimensional inward-turning inlet.

[0013] Furthermore, in S2, the method for preprocessing the data sets corresponding to different Mach numbers is: sampling processing, time alignment processing, spatial alignment processing and standardization processing on the data sets; wherein, the Mach number levels include 4.0Ma, 5.0Ma, 6.0Ma and 7.0Ma.

[0014] Furthermore, the inward-turning inlet flow field reconstruction model includes a first input layer, a first fully connected layer, a first dimensionality transformation layer, a first convolutional layer, a first normalization layer, a first maximum pooling layer, a first LIF layer, a second convolutional layer, a second normalization layer, a second LIF layer, a third convolutional layer, a third normalization layer, a third LIF layer, a fourth convolutional layer, a fourth normalization layer, an attention mechanism unit, a first adder, a fourth LIF layer, a second dimensionality transformation layer, a fourth fully connected layer and a first output layer;

[0015] The input end of the first input layer serves as the input end of the inward-turning inlet flow field reconstruction model, and its output end is connected to the input end of the first fully connected layer; the output end of the first fully connected layer, the output end of the first dimensionality transformation layer, the first convolutional layer, the first normalization layer, the first maximum pooling layer, the first LIF layer, the second convolutional layer and the input end of the second normalization layer are connected in sequence; the first output end of the second normalization layer, the second LIF layer, the third convolutional layer, the third normalization layer, the third LIF layer, the fourth convolutional layer, the fourth normalization layer and the input end of the attention mechanism unit are connected in sequence; the output end of the attention mechanism unit and the second output end of the second normalization layer are both connected to the input end of the first adder; the output end of the first adder, the fourth LIF layer, the second dimensionality transformation layer, the fourth fully connected layer and the input end of the first output layer are connected in sequence; the output end of the first output layer serves as the output end of the inward-turning inlet flow field reconstruction model.

[0016] Furthermore, the attention mechanism unit includes a channel attention module and a spatial attention module connected in sequence;

[0017] The channel attention module includes a second input layer, a second maximum pooling layer, a first global pooling layer, a second fully connected layer, a first activation function layer, a third fully connected layer, a second adder, a fifth normalization layer, a third dimension transformation layer and a first multiplier;

[0018] The spatial attention module includes the third maximum pooling layer, the second global pooling layer, the first feature fusion layer, the fifth convolutional layer, the sixth normalization layer, the second multiplier and the second output layer;

[0019] The input end of the second input layer serves as the input end of the channel attention module, its first output end is connected to the input end of the second maximum pooling layer, and its second output end is connected to the input end of the first global pooling layer; the output end of the second maximum pooling layer is connected to the first input end of the second fully connected layer; the output end of the first global pooling layer is connected to the second input end of the second fully connected layer; the first output end of the second fully connected layer is connected to the first input end of the first activation function layer; the second output end of the second fully connected layer is connected to the second input end of the first activation function layer; the first output end of the first activation function layer is connected to the first input end of the third fully connected layer; the second output end of the first activation function layer is connected to the second input end of the third fully connected layer; the first output end of the third fully connected layer is connected to the first input end of the second adder; the second output end of the third fully connected layer is connected to the second input end of the second adder; the output end of the second adder is connected to the input end of the fifth normalization layer The output of the fifth normalization layer is connected to the input of the third dimensional change layer; the output of the third dimensional transformation layer is connected to the first input of the first multiplier; the third output of the second input layer is connected to the second input of the first multiplier; the first output of the first multiplier is connected to the input of the third maximum pooling layer, and its second output is connected to the input of the second global pooling layer; the output of the third maximum pooling layer is connected to the first input of the first feature fusion layer; the output of the second global pooling layer is connected to the second input of the first feature fusion layer; the output of the first feature fusion layer, the fifth convolutional layer and the input of the sixth normalization layer are connected in sequence; the output of the sixth normalization layer is connected to the first input of the second multiplier; the third output of the first multiplier is connected to the second input of the second multiplier; the output of the second multiplier is connected to the input of the second output layer; the output of the second output layer serves as the output of the spatial attention module.

[0020] Furthermore, the expression of the channel attention mechanism module is:

[0021]

[0022]

[0023] Where, represents the output of the channel attention mechanism module, g c (·) represents a one-dimensional weight function, U t,n represents the input of the channel attention mechanism module, σ(·) represents the sigmoid function, represents the first multilayer perceptron, Represents the second multilayer perceptron, ReLU(·) represents the activation function, AvgPool(·) represents the average pooling operation, and MaxPool(·) represents the maximum pooling operation.

[0024] Furthermore, the expression of the spatial attention mechanism module is:

[0025]

[0026] g s (U t,n )=σ(f 3×3 ([AvgPool(U t,n );MaxPool(U t,n )]))

[0027] Where, represents the output of the spatial attention mechanism unit, g s (·) represents the two-dimensional SA attention weight function, U t ,n represents the input of the channel attention mechanism module, f 3×3 represents a 3×3 convolution operation with filter size, AvgPool(·) represents an average pooling operation, and MaxPool(·) represents a maximum pooling operation.

[0028] Furthermore, the shock wave boundary layer separation point position detection model includes a third input layer, a fifth fully connected layer, a fourth dimensional transformation layer, a sixth convolutional layer, a second activation function layer, a seventh normalization layer, a seventh convolutional layer, a third activation function layer, an eighth normalization layer, an eighth convolutional layer, a fourth activation function layer, a ninth normalization layer, a third adder, a fourth maximum pooling layer, a ninth convolutional layer, a tenth normalization layer, a fifth activation function layer, a tenth convolutional layer, an eleventh normalization layer, a sixth activation function layer, a fifth maximum pooling layer, an eleventh convolutional layer, a twelfth normalization layer, a seventh activation function layer, a twelfth convolutional layer, a thirteenth normalization layer, an eighth activation function layer, a second feature fusion layer, a sixth fully connected layer and a fourth output layer;

[0029] The input end of the third input layer serves as the input end of the shock wave boundary layer separation point position detection model; the output end of the third input layer, the fifth fully connected layer, the fourth dimensional transformation layer, the sixth convolutional layer, the second activation function layer and the input end of the seventh normalization layer are connected in sequence; the first output end of the seventh normalization layer, the seventh convolutional layer, the third activation function layer, the eighth normalization layer, the eighth convolutional layer, the fourth activation function layer and the input end of the ninth normalization layer are connected in sequence; the output end of the ninth normalization layer is connected to the first input end of the third adder; the second output end of the seventh normalization layer is connected to the second input end of the third adder; the first output end of the third adder is connected to the input end of the fourth maximum pooling layer; the output end of the fourth maximum pooling layer, the ninth convolutional layer, the tenth normalization layer , the input ends of the fifth activation function layer, the tenth convolution layer, the eleventh normalization layer and the sixth activation function layer are connected in sequence; the second output end of the third adder is connected to the input end of the fifth maximum pooling layer; the output end of the fifth maximum pooling layer, the eleventh convolution layer, the twelfth normalization layer, the seventh activation function layer, the twelfth convolution layer, the thirteenth normalization layer and the input end of the eighth activation function layer are connected in sequence; the output end of the sixth activation function layer is connected to the first input end of the second feature fusion layer; the output end of the eighth activation function layer is connected to the second input end of the second feature fusion layer; the output end of the second feature fusion layer, the sixth fully connected layer and the input end of the fourth output layer are connected in sequence; the output end of the fourth output layer serves as the output end of the shock boundary layer separation point position detection model.

[0030] The beneficial effects of the present invention are as follows: the present invention utilizes a convolutional neural network for feature extraction, establishes a nonlinear feature mapping relationship between input data and output flow fields, and introduces pulses into the convolutional neural network, effectively solving the problem of high computational cost of the original model; not only does it ensure prediction accuracy, but it also has the advantages of low latency and low power consumption at the hardware level. At the same time, the present invention utilizes a convolutional neural network to construct a shock wave boundary layer separation point position detection model for the unsteady flow field of an inward-turning inlet, quickly locates the shock wave boundary layer separation zone and makes a decision, and finally utilizes an FPGA platform to realize visualization of the flow field and shock wave boundary layer separation point position, which is of great significance for improving the starting capability of the inlet, promoting the optimization of the inlet design, and ensuring the stable operation of the engine.

[0031] Based on the above method, the present invention also proposes an intelligent reconstruction system for the unsteady flow field of a wide-range internal rotation inlet, which includes a data set generation unit, a data set preprocessing unit, a flow field prediction unit, a separation point data set generation unit, a separation point prediction unit, and a visualization display unit;

[0032] The data set generation unit is used to perform CFD numerical simulation on the two-dimensional inward-turning inlet at different Mach numbers to obtain data sets corresponding to different Mach numbers;

[0033] The data set preprocessing unit is used to preprocess the data sets corresponding to different Mach numbers;

[0034] The flow field prediction unit is used to construct a flow field reconstruction model for the inward-turning inlet, input the pre-processed data sets corresponding to different Mach numbers into the inward-turning inlet flow field reconstruction model, and output the predicted flow field;

[0035] The separation point data set generation unit is used to obtain the distance from the shock wave boundary layer separation point to the two-dimensional inward-turning inlet inlet, and generate a shock wave boundary layer separation point position detection data set;

[0036] The separation point prediction unit is used to construct a shock wave boundary layer separation point position detection model, input the shock wave boundary layer separation point position detection data set into the shock wave boundary layer separation point position detection model, and output the predicted shock wave boundary layer separation point position;

[0037] The visualization display unit is used to visualize the predicted flow field and the predicted shock wave boundary layer separation points.

[0038] The beneficial effect of the present invention is that the system can realize the visualization of the flow field and the position of the shock wave boundary layer separation point, which is of great significance for improving the starting capability of the intake duct, promoting the optimization of the intake duct design and ensuring the stable operation of the engine. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of the intelligent reconstruction method for the unsteady flow field of a wide-area internal rotating inlet;

[0040] Figure 2 This is a schematic diagram of the structure of the inward-turning inlet flow field reconstruction model;

[0041] Figure 3 Schematic diagram of the structure of the attention mechanism unit;

[0042] Figure 4 This is a schematic diagram of the structure of the shock wave boundary layer separation point detection model;

[0043] Figure 5 Schematic diagram of the structure of the intelligent reconstruction system for the unsteady flow field of a wide-area internal rotating inlet. DETAILED DESCRIPTION

[0044] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0045] like Figure 1 As shown, the present invention provides a method for intelligent reconstruction of unsteady flow field of a wide-range internal rotation inlet, comprising the following steps:

[0046] S1. Perform CFD numerical simulation on the two-dimensional inward-turning inlet at different Mach numbers to obtain data sets corresponding to different Mach numbers.

[0047] S2. Preprocessing the data sets corresponding to different Mach numbers;

[0048] S3. Constructing an inward-turning inlet flow field reconstruction model, inputting the pre-processed data sets corresponding to different Mach numbers into the inward-turning inlet flow field reconstruction model, and outputting a predicted flow field;

[0049] S4. Obtain the distance from the shock wave boundary layer separation point to the two-dimensional inward-turning inlet inlet, and generate a shock wave boundary layer separation point position detection data set;

[0050] S5. Constructing a shock wave boundary layer separation point position detection model, inputting the shock wave boundary layer separation point position detection data set into the shock wave boundary layer separation point position detection model, and outputting a predicted shock wave boundary layer separation point position;

[0051] S6. Visualize the predicted flow field and the predicted shock wave boundary layer separation points.

[0052] In an embodiment of the present invention, in S1, the data sets corresponding to different Mach numbers include upper and lower wall pressure data, pressure field, density field, schlieren image, velocity field in the x-direction, and velocity field in the y-direction of the two-dimensional inward-turning inlet.

[0053] In the embodiment of the present invention, the entire configuration is an inward-turning air intake represented by a smooth Bezier curve and defined by seven control points. The control points of the Bezier curve determine the air intake shape through 12 design parameters: the magnitude vector of the leading edge angle, the direction θ le , intermediate control point (x m , r m ), control vector λ m1 and λ m2 The magnitude vector and direction θ m1 and θ m2 , the size curve λ of the inlet outlet c , direction θ c , Combustion chamber radius r c and the inlet radius r i . Air r i The inlet radius is fixed at 0.1 m to obtain a constant mass flow capture rate, and an appropriate grid resolution is set to ensure that the grid size does not affect the prediction of the inlet flow field.

[0054] In the CFD numerical simulation software, the initial inlet flow conditions at the inlet were set to Mach altitude (km) = 33.4, static temperature (K) T = 232, static pressure (Pa) P = 724, grid number = 32,000, and Mach numbers (Ma) = 4.0, 5.0, 6.0, 7.0, and 8.0. Data were acquired from t = 0.02 ms to t = 0.2 ms, with a step length of 0.02 ms. The numerical simulations collected 500 sets of data, including upper and lower wall pressure data, pressure field (p), density field (r), schlieren images, and velocity fields in all directions, including the x-direction velocity field (u) and the y-direction velocity field (v).

[0055] In an embodiment of the present invention, in S2, the method for preprocessing the data sets corresponding to different Mach numbers is: performing sampling processing, time alignment processing, spatial alignment processing and standardization processing on the data sets; wherein the Mach number levels include 4.0Ma, 5.0Ma, 6.0Ma and 7.0Ma.

[0056] In this embodiment of the present invention, to better extract features and predict unsteady flow fields, the acquired data is aligned in time and space, and each pair of sample data (a pair of samples includes upper and lower wall pressure, Mach number, time, and the corresponding flow field) is individually normalized. Data with a Mach number of 8.0 can also be used as a test set for subsequent testing of the inward-turning inlet flow field reconstruction model.

[0057] In the embodiment of the present invention, Figure 2 As shown, the inward-turning inlet flow field reconstruction model includes a first input layer, a first fully connected layer, a first dimensionality transformation layer, a first convolutional layer, a first normalization layer, a first maximum pooling layer, a first LIF layer, a second convolutional layer, a second normalization layer, a second LIF layer, a third convolutional layer, a third normalization layer, a third LIF layer, a fourth convolutional layer, a fourth normalization layer, an attention mechanism unit, a first adder, a fourth LIF layer, a second dimensionality transformation layer, a fourth fully connected layer and a first output layer;

[0058] The input end of the first input layer serves as the input end of the inward-turning inlet flow field reconstruction model, and its output end is connected to the input end of the first fully connected layer; the output end of the first fully connected layer, the output end of the first dimensionality transformation layer, the first convolutional layer, the first normalization layer, the first maximum pooling layer, the first LIF layer, the second convolutional layer and the input end of the second normalization layer are connected in sequence; the first output end of the second normalization layer, the second LIF layer, the third convolutional layer, the third normalization layer, the third LIF layer, the fourth convolutional layer, the fourth normalization layer and the input end of the attention mechanism unit are connected in sequence; the output end of the attention mechanism unit and the second output end of the second normalization layer are both connected to the input end of the first adder; the output end of the first adder, the fourth LIF layer, the second dimensionality transformation layer, the fourth fully connected layer and the input end of the first output layer are connected in sequence; the output end of the first output layer serves as the output end of the inward-turning inlet flow field reconstruction model.

[0059] In an embodiment of the present invention, the model is a pulse residual neural network based on an attention mechanism. First, a fully connected layer is used to change the dimension to facilitate subsequent convolution operations, and then the time dimension in the pulse neural network is increased through dimensional transformation, and converted into a form that can be processed by the neural network. After that, operations such as 1*1 convolution, normalization and pooling are performed to change the number of feature channels, help fuse different features, and accelerate training. The LIF model with a simple form and low computational complexity is selected as the pulse neuron. At the same time, since attention optimizes the membrane potential in a data-dependent manner, it can lead to fewer pulse responses and produce better performance and energy efficiency, and can solve the degradation problem of general deep pulse neural networks. Therefore, the attention mechanism is introduced to realize the feature output of the convolution layer from the two dimensions of channel and space to obtain higher-dimensional features. Finally, the dimension is transformed again to reduce the data dimension and calculate the mean of the time dimension, which is processed through the fully connected layer and then used as the network output.

[0060] The model uses a pulse residual neural network based on the attention mechanism to establish a nonlinear feature mapping relationship between the input and the corresponding flow field. Its expression is: Y = S(θ, x, y, t, ma); where S represents the pulse residual neural network, Y represents the predicted flow field, θ represents the model parameters of the pulse residual neural network, x is the upper wall pressure, y is the lower wall pressure, t represents time, and ma represents the Mach number.

[0061] The attention mechanism of this model includes channel attention mechanism and spatial attention mechanism, which are composed of these two mechanism modules in series. The input feature map first passes through the channel attention mechanism, and the channel weight is multiplied by the input feature map before being sent to the spatial attention mechanism. The normalized spatial weight is multiplied by the input feature map of the spatial attention mechanism to obtain the final weighted feature map. The channel attention mechanism uses the function g c(·) to directly refine the membrane potential of spiking neurons, suppressing secondary features and improving efficiency.

[0062] The loss function uses MSELoss, and its calculation formula is: In the formula, n represents the number of training times of the neural network model, x i represents the flow field obtained by numerical calculation of the i-th sample, y i represents the flow field predicted by the pulse residual neural network output of the i-th sample.

[0063] In the embodiment of the present invention, Figure 3 As shown, the attention mechanism unit includes a channel attention module and a spatial attention module connected in sequence;

[0064] The channel attention module includes a second input layer, a second maximum pooling layer, a first global pooling layer, a second fully connected layer, a first activation function layer, a third fully connected layer, a second adder, a fifth normalization layer, a third dimension transformation layer and a first multiplier;

[0065] The spatial attention module includes the third maximum pooling layer, the second global pooling layer, the first feature fusion layer, the fifth convolutional layer, the sixth normalization layer, the second multiplier and the second output layer;

[0066] The input end of the second input layer serves as the input end of the channel attention module, its first output end is connected to the input end of the second maximum pooling layer, and its second output end is connected to the input end of the first global pooling layer; the output end of the second maximum pooling layer is connected to the first input end of the second fully connected layer; the output end of the first global pooling layer is connected to the second input end of the second fully connected layer; the first output end of the second fully connected layer is connected to the first input end of the first activation function layer; the second output end of the second fully connected layer is connected to the second input end of the first activation function layer; the first output end of the first activation function layer is connected to the first input end of the third fully connected layer; the second output end of the first activation function layer is connected to the second input end of the third fully connected layer; the first output end of the third fully connected layer is connected to the first input end of the second adder; the second output end of the third fully connected layer is connected to the second input end of the second adder; the output end of the second adder is connected to the input end of the fifth normalization layer The output of the fifth normalization layer is connected to the input of the third dimensional change layer; the output of the third dimensional transformation layer is connected to the first input of the first multiplier; the third output of the second input layer is connected to the second input of the first multiplier; the first output of the first multiplier is connected to the input of the third maximum pooling layer, and its second output is connected to the input of the second global pooling layer; the output of the third maximum pooling layer is connected to the first input of the first feature fusion layer; the output of the second global pooling layer is connected to the second input of the first feature fusion layer; the output of the first feature fusion layer, the fifth convolutional layer and the input of the sixth normalization layer are connected in sequence; the output of the sixth normalization layer is connected to the first input of the second multiplier; the third output of the first multiplier is connected to the second input of the second multiplier; the output of the second multiplier is connected to the input of the second output layer; the output of the second output layer serves as the output of the spatial attention module.

[0067] In this embodiment of the present invention, the expression of the channel attention mechanism module is:

[0068]

[0069]

[0070] Where, represents the output of the channel attention mechanism module, g c (·) represents a one-dimensional weight function, U t,n represents the input of the channel attention mechanism module, σ(·) represents the sigmoid function, represents the first multilayer perceptron, Represents the second multilayer perceptron, ReLU(·) represents the activation function, AvgPool(·) represents the average pooling operation, and MaxPool(·) represents the maximum pooling operation.

[0071] In this embodiment of the present invention, the expression of the spatial attention mechanism module is:

[0072]

[0073] g s (U t,n )=σ(f 3×3 ([AvgPool(U t,n );MaxPool(U t,n )]))

[0074] Where, represents the output of the spatial attention mechanism unit, g s (·) represents the two-dimensional SA attention weight function, U t ,n represents the input of the channel attention mechanism module, f 3×3 represents a 3×3 convolution operation with filter size, AvgPool(·) represents an average pooling operation, and MaxPool(·) represents a maximum pooling operation.

[0075] In the embodiment of the present invention, Figure 4 As shown, the shock wave boundary layer separation point position detection model includes a third input layer, a fifth fully connected layer, a fourth dimensional transformation layer, a sixth convolutional layer, a second activation function layer, a seventh normalization layer, a seventh convolutional layer, a third activation function layer, an eighth normalization layer, an eighth convolutional layer, a fourth activation function layer, a ninth normalization layer, a third adder, a fourth maximum pooling layer, a ninth convolutional layer, a tenth normalization layer, a fifth activation function layer, a tenth convolutional layer, an eleventh normalization layer, a sixth activation function layer, a fifth maximum pooling layer, an eleventh convolutional layer, a twelfth normalization layer, a seventh activation function layer, a twelfth convolutional layer, a thirteenth normalization layer, an eighth activation function layer, a second feature fusion layer, a sixth fully connected layer and a fourth output layer;

[0076] The input end of the third input layer serves as the input end of the shock wave boundary layer separation point position detection model; the output end of the third input layer, the fifth fully connected layer, the fourth dimensional transformation layer, the sixth convolutional layer, the second activation function layer and the input end of the seventh normalization layer are connected in sequence; the first output end of the seventh normalization layer, the seventh convolutional layer, the third activation function layer, the eighth normalization layer, the eighth convolutional layer, the fourth activation function layer and the input end of the ninth normalization layer are connected in sequence; the output end of the ninth normalization layer is connected to the first input end of the third adder; the second output end of the seventh normalization layer is connected to the second input end of the third adder; the first output end of the third adder is connected to the input end of the fourth maximum pooling layer; the output end of the fourth maximum pooling layer, the ninth convolutional layer, the tenth normalization layer , the input ends of the fifth activation function layer, the tenth convolution layer, the eleventh normalization layer and the sixth activation function layer are connected in sequence; the second output end of the third adder is connected to the input end of the fifth maximum pooling layer; the output end of the fifth maximum pooling layer, the eleventh convolution layer, the twelfth normalization layer, the seventh activation function layer, the twelfth convolution layer, the thirteenth normalization layer and the input end of the eighth activation function layer are connected in sequence; the output end of the sixth activation function layer is connected to the first input end of the second feature fusion layer; the output end of the eighth activation function layer is connected to the second input end of the second feature fusion layer; the output end of the second feature fusion layer, the sixth fully connected layer and the input end of the fourth output layer are connected in sequence; the output end of the fourth output layer serves as the output end of the shock boundary layer separation point position detection model.

[0077] In an embodiment of the present invention, the model is mainly composed of operations such as a fully connected layer, a convolution, a maximum pooling layer, and a residual connection. In order to enhance the model's ability to extract features, the dimension is first transformed through the fully connected layer and the Reshape() function of pytorch. Then, through the residual connection, the shallow low-frequency features and the deep high-frequency features are fused pixel by pixel, which can reduce the loss of features during the convolution process and protect the integrity of the information. Then, a conventional convolution layer is used to extract features and reduce the size. In order to extract different types of features and improve the performance of the model, different feature maps are extracted through two branches and fused using the Concat() function of pytorch. Finally, the final shock wave boundary layer separation point position is obtained through the fully connected layer.

[0078] The model uses a convolutional neural network to establish a nonlinear feature mapping relationship between the predicted schlieren image and the location of the shock boundary layer separation point. Its expression is: D = F(θ, p); where F(·) represents the mapping function, D represents the location of the shock boundary layer separation point, θ represents all learnable parameters of the convolutional neural network, and p represents the predicted schlieren image.

[0079] The beneficial effects of the present invention are as follows: in the reconstruction of the unsteady flow field of the inward-turning inlet, a convolutional neural network is used to extract features, a nonlinear feature mapping relationship between input data and output flow field is established, and pulses are introduced into the convolutional neural network, which effectively solves the problem of high computational cost of the original model. It not only ensures the prediction accuracy, but also has the advantages of low latency and low power consumption at the hardware level. At the same time, a convolutional neural network is used to construct a shock wave boundary layer separation point position detection model for the unsteady flow field of the inward-turning inlet, quickly locate the shock wave boundary layer separation area and make a decision, and finally use the FPGA platform to realize the visualization of the flow field and shock wave boundary layer separation point position and transmit it to the controller. This is of great significance for improving the starting capability of the inlet, promoting the optimization of the inlet design, and ensuring the stable operation of the engine.

[0080] Based on the above method, the present invention also proposes a wide-range internal rotation inlet unsteady flow field intelligent reconstruction system, such as Figure 5 As shown, it includes a data set generation unit, a data set preprocessing unit, a flow field prediction unit, a separation point data set generation unit, a separation point prediction unit and a visualization display unit;

[0081] The data set generation unit is used to perform CFD numerical simulation on the two-dimensional inward-turning inlet at different Mach numbers to obtain data sets corresponding to different Mach numbers;

[0082] The data set preprocessing unit is used to preprocess the data sets corresponding to different Mach numbers;

[0083] The flow field prediction unit is used to construct a flow field reconstruction model for the inward-turning inlet, input the pre-processed data sets corresponding to different Mach numbers into the inward-turning inlet flow field reconstruction model, and output the predicted flow field;

[0084] The separation point data set generation unit is used to obtain the distance from the shock wave boundary layer separation point to the two-dimensional inward-turning inlet inlet, and generate a shock wave boundary layer separation point position detection data set;

[0085] The separation point prediction unit is used to construct a shock wave boundary layer separation point position detection model, input the shock wave boundary layer separation point position detection data set into the shock wave boundary layer separation point position detection model, and output the predicted shock wave boundary layer separation point position;

[0086] The visualization display unit is used to visualize the predicted flow field and the predicted shock wave boundary layer separation points.

[0087] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. An intelligent reconstruction method for unsteady flow field of wide-range internal rotation inlet, characterized by: The following steps are involved: S1. Perform CFD numerical simulation on the two-dimensional inward-turning inlet at different Mach numbers to obtain data sets corresponding to different Mach numbers. S2. Preprocessing the data sets corresponding to different Mach numbers; S3. Constructing an inward-turning inlet flow field reconstruction model, inputting the pre-processed data sets corresponding to different Mach numbers into the inward-turning inlet flow field reconstruction model, and outputting a predicted flow field; S4. Obtain the distance from the shock wave boundary layer separation point to the two-dimensional inward-turning inlet inlet, and generate a shock wave boundary layer separation point position detection data set; S5. Constructing a shock wave boundary layer separation point position detection model, inputting the shock wave boundary layer separation point position detection data set into the shock wave boundary layer separation point position detection model, and outputting a predicted shock wave boundary layer separation point position; S6. Visualize the predicted flow field and predicted shock wave boundary layer separation points; The inward-turning inlet flow field reconstruction model uses a pulse residual neural network based on the attention mechanism to establish a nonlinear feature mapping relationship between the input and the corresponding flow field, which is expressed as follows: Y = S ( θ , x , y , t , ma ); where S represents the pulse residual neural network, Y represents the predicted flow field, θ represents the model parameters of the pulse residual neural network, x is the upper wall pressure, y is the lower wall pressure, t Indicates time, ma Represents the Mach number.

2. The method for intelligent reconstruction of unsteady flow field of wide-range internal rotation inlet according to claim 1 is characterized in that: In said S1, the data sets corresponding to different Mach numbers include upper and lower wall pressure data, pressure field, density field, schlieren image, x The velocity field in the direction and y Velocity field in the direction.

3. The method for intelligent reconstruction of unsteady flow field of wide-area internal rotation inlet according to claim 1 is characterized in that: In S2, the method for preprocessing the data sets corresponding to different Mach numbers is: performing sampling processing, time alignment processing, spatial alignment processing and standardization processing on the data sets; wherein the Mach number levels include 4.0Ma, 5.0Ma, 6.0Ma and 7.0Ma.

4. The method for intelligent reconstruction of unsteady flow field of wide-range internal rotation inlet according to claim 1 is characterized in that: The inner-rotating inlet flow field reconstruction model includes a first input layer, a first fully connected layer, a first dimensionality transformation layer, a first convolutional layer, a first normalization layer, a first maximum pooling layer, a first LIF layer, a second convolutional layer, a second normalization layer, a second LIF layer, a third convolutional layer, a third normalization layer, a third LIF layer, a fourth convolutional layer, a fourth normalization layer, an attention mechanism unit, a first adder, a fourth LIF layer, a second dimensionality transformation layer, a fourth fully connected layer and a first output layer; The input end of the first input layer serves as the input end of the inward-turning air inlet flow field reconstruction model, and its output end is connected to the input end of the first fully connected layer; the output end of the first fully connected layer, the output end of the first dimensionality transformation layer, the first convolutional layer, the first normalization layer, the first maximum pooling layer, the first LIF layer, the second convolutional layer and the input end of the second normalization layer are connected in sequence; the first output end of the second normalization layer, the second LIF layer, the third convolutional layer, the third normalization layer, the third LIF layer, the fourth convolutional layer, the fourth normalization layer and the input end of the attention mechanism unit are connected in sequence; the output end of the attention mechanism unit and the second output end of the second normalization layer are both connected to the input end of the first adder; the output end of the first adder, the fourth LIF layer, the second dimensionality transformation layer, the fourth fully connected layer and the input end of the first output layer are connected in sequence; the output end of the first output layer serves as the output end of the inward-turning air inlet flow field reconstruction model.

5. The method for intelligent reconstruction of unsteady flow field of wide-area internal rotation inlet according to claim 4 is characterized in that: The attention mechanism unit includes a channel attention module and a spatial attention module connected in sequence; The channel attention module includes a second input layer, a second maximum pooling layer, a first global pooling layer, a second fully connected layer, a first activation function layer, a third fully connected layer, a second adder, a fifth normalization layer, a third dimension transformation layer and a first multiplier; The spatial attention module includes a third maximum pooling layer, a second global pooling layer, a first feature fusion layer, a fifth convolutional layer, a sixth normalization layer, a second multiplier and a second output layer; The input end of the second input layer serves as the input end of the channel attention module, its first output end is connected to the input end of the second maximum pooling layer, and its second output end is connected to the input end of the first global pooling layer; the output end of the second maximum pooling layer is connected to the first input end of the second fully connected layer; the output end of the first global pooling layer is connected to the second input end of the second fully connected layer; the first output end of the second fully connected layer is connected to the first input end of the first activation function layer; the second output end of the second fully connected layer is connected to the second input end of the first activation function layer; the first output end of the first activation function layer is connected to the first input end of the third fully connected layer; the second output end of the first activation function layer is connected to the second input end of the third fully connected layer; the first output end of the third fully connected layer is connected to the first input end of the second adder; the second output end of the third fully connected layer and The second input end of the second adder is connected; the output end of the second adder is connected to the input end of the fifth normalization layer; the output end of the fifth normalization layer is connected to the input end of the third dimensional change layer; the output end of the third dimensional transformation layer is connected to the first input end of the first multiplier; the third output end of the second input layer is connected to the second input end of the first multiplier; the first output end of the first multiplier is connected to the input end of the third maximum pooling layer, and the second output end of the first multiplier is connected to the input end of the second global pooling layer; the output end of the third maximum pooling layer is connected to the first input end of the first feature fusion layer; the output end of the second global pooling layer is connected to the second input end of the first feature fusion layer; the output end of the first feature fusion layer, the fifth convolutional layer and the input end of the sixth normalization layer are connected in sequence; the output end of the sixth normalization layer is connected to the first input end of the second multiplier; The third output terminal of the first multiplier is connected to the second input terminal of the second multiplier; the output terminal of the second multiplier is connected to the input terminal of the second output layer; and the output terminal of the second output layer serves as the output terminal of the spatial attention module.

6. The method for intelligent reconstruction of unsteady flow field of wide-area internal rotation inlet according to claim 5 is characterized in that: The expression of the channel attention module is: Where, represents the output of the channel attention module, represents a one-dimensional weight function, represents the input of the channel attention module, express sigmoid function, represents the first multilayer perceptron, represents the second multilayer perceptron, represents the activation function, represents the average pooling operation, Represents the maximum pooling operation.

7. The method for intelligent reconstruction of unsteady flow field of wide-area internal rotation inlet according to claim 5 is characterized in that: The expression of the spatial attention module is: Where, represents the output of the spatial attention unit, represents the two-dimensional SA attention weight function, represents the input of the channel attention module, represents a 3×3 convolution operation with filter size, represents the average pooling operation, Represents the maximum pooling operation.

8. The method for intelligent reconstruction of unsteady flow field of wide-area internal rotation inlet according to claim 1 is characterized in that: The shock wave boundary layer separation point position detection model includes a third input layer, a fifth fully connected layer, a fourth dimensional transformation layer, a sixth convolutional layer, a second activation function layer, a seventh normalization layer, a seventh convolutional layer, a third activation function layer, an eighth normalization layer, an eighth convolutional layer, a fourth activation function layer, a ninth normalization layer, a third adder, a fourth maximum pooling layer, a ninth convolutional layer, a tenth normalization layer, a fifth activation function layer, a tenth convolutional layer, an eleventh normalization layer, a sixth activation function layer, a fifth maximum pooling layer, an eleventh convolutional layer, a twelfth normalization layer, a seventh activation function layer, a twelfth convolutional layer, a thirteenth normalization layer, an eighth activation function layer, a second feature fusion layer, a sixth fully connected layer and a fourth output layer; The input end of the third input layer serves as the input end of the shock boundary layer separation point position detection model; the output end of the third input layer, the fifth fully connected layer, the fourth dimensional transformation layer, the sixth convolutional layer, the second activation function layer, and the input end of the seventh normalization layer are connected in sequence; the first output end of the seventh normalization layer, the seventh convolutional layer, the third activation function layer, the eighth normalization layer, the eighth convolutional layer, the fourth activation function layer, and the input end of the ninth normalization layer are connected in sequence; The output end of the ninth normalization layer is connected to the first input end of the third adder; the second output end of the seventh normalization layer is connected to the second input end of the third adder; the first output end of the third adder is connected to the input end of the fourth maximum pooling layer; The output end of the fourth maximum pooling layer, the ninth convolutional layer, the tenth normalization layer, the fifth activation function layer, the tenth convolutional layer, the eleventh normalization layer, and the input end of the sixth activation function layer are connected in sequence; the second output end of the third adder is connected to the input end of the fifth maximum pooling layer; The output end of the fifth maximum pooling layer, the eleventh convolution layer, the twelfth normalization layer, the seventh activation function layer, the twelfth convolution layer, the thirteenth normalization layer and the input end of the eighth activation function layer are connected in sequence; the output end of the sixth activation function layer is connected to the first input end of the second feature fusion layer; the output end of the eighth activation function layer is connected to the second input end of the second feature fusion layer; the output end of the second feature fusion layer, the sixth fully connected layer and the input end of the fourth output layer are connected in sequence; the output end of the fourth output layer serves as the output end of the shock boundary layer separation point position detection model.

9. An intelligent reconstruction system for unsteady flow field of wide-range internal rotation inlet, characterized by: The wide-area internal rotation inlet unsteady flow field intelligent reconstruction system is implemented using a wide-area internal rotation inlet unsteady flow field intelligent reconstruction method. The system includes a data set generation unit, a data set preprocessing unit, a flow field prediction unit, a separation point data set generation unit, a separation point prediction unit, and a visualization display unit. The data set generating unit is used to perform CFD numerical simulation on the two-dimensional inward-turning inlet at different Mach numbers to obtain data sets corresponding to different Mach numbers; The data set preprocessing unit is used to preprocess data sets corresponding to different Mach numbers; The flow field prediction unit is used to construct an inward-turning inlet flow field reconstruction model, and input the pre-processed data sets corresponding to different Mach numbers into the inward-turning inlet flow field reconstruction model to output a predicted flow field; The separation point data set generation unit is used to obtain the distance from the shock wave boundary layer separation point to the two-dimensional inward-turning air inlet inlet, and generate a shock wave boundary layer separation point position detection data set; The separation point prediction unit is used to construct a shock wave boundary layer separation point position detection model, input the shock wave boundary layer separation point position detection data set into the shock wave boundary layer separation point position detection model, and output a predicted shock wave boundary layer separation point position; The visualization display unit is used to visualize the predicted flow field and the predicted shock wave boundary layer separation point; The inward-turning inlet flow field reconstruction model uses a pulse residual neural network based on the attention mechanism to establish a nonlinear feature mapping relationship between the input and the corresponding flow field, which is expressed as follows: Y = S ( θ , x , y , t , ma ); where S represents the pulse residual neural network, Y represents the predicted flow field, θ represents the model parameters of the pulse residual neural network, x is the upper wall pressure, y is the lower wall pressure, t Indicates time, ma Represents the Mach number.

Images