External store release separation trajectory prediction method based on small sample size

Abstract

The invention relates to the technical field of aircraft external store release separation trajectory prediction, in particular to an external store release separation trajectory prediction method based on a small sample size. The external store aerodynamic force prediction model and the trajectory prediction model are established, and an iterative prediction method is adopted, so that the dependence on an external store release separation data set is effectively reduced, and the prediction precision and efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft external attachment release and separation trajectory prediction, and in particular to a method for predicting the release and separation trajectory of external attachments based on a small sample size. Background Art

[0002] During the separation process of external stores from an aircraft, accurate calculation of the trajectory of the external stores is crucial for flight safety and mission completion. Traditional methods for obtaining the trajectory of external stores separation rely on numerical simulation methods based on computational fluid dynamics (CFD) or are obtained through wind tunnel tests. With the development of machine learning-related algorithms and technologies, machine learning methods represented by neural networks have begun to be applied to the prediction of external stores release and separation. For example, reference 1 (Prediction of external stores release and separation based on feedforward neural networks, Journal of Ballistics, 2023.09.) uses feedforward neural networks to accurately model and predict the trajectory and posture of external stores during the release and separation movement. Conventional external stores release and separation trajectory prediction methods usually rely on a large number of data sets to train neural networks to improve the prediction accuracy of the network. However, in practical applications, obtaining a large number of high-quality external stores release and separation data sets is often costly and time-consuming. Specifically, the following aspects limit the application of traditional methods. First, data acquisition is difficult. The separation process of aircraft external stores release involves complex aerodynamic environments and variable flight conditions, and it is costly to obtain a large amount of high-quality experimental data or simulation data. Second, the measurement of experimental data is complex. The test of the external storage release separation trajectory requires high-precision measurement equipment and professional technicians, and the data measurement process is complex and time-consuming. The third is data sparsity. Under certain specific flight conditions, the external storage release separation data may be very sparse, which is difficult to meet the needs of traditional neural network training.

[0003] Therefore, how to accurately predict the trajectory of the external attachment release and separation under the condition of limited data has become a technical problem that needs to be solved urgently. The traditional external attachment trajectory prediction method based on neural network only uses the external attachment's appearance characteristics as the input of the network, and requires a large amount of data sets to establish a nonlinear mapping between the external attachment's appearance characteristics and the separation trajectory. In fact, during the release and separation process, there is a connection between the aerodynamic history of the external attachment, that is, the aerodynamic component of the external attachment at each separation time step, and the trajectory history of the external attachment, that is, the center of gravity position component of the external attachment at each separation time step. The position of the external attachment determines the flow field distribution of the external attachment, thereby affecting the aerodynamic force received by the external attachment. The aerodynamic force of the external attachment determines the movement of the external attachment, thereby affecting the position of the external attachment, and the two are coupled with each other. According to this coupling phenomenon, the aerodynamic history of the external attachment can be used as a supplementary input of the neural network to improve the input feature dimension of the neural network and achieve high-precision separation trajectory prediction under the condition of limited data. Summary of the invention

[0004] The purpose of the present invention is to propose a method for predicting the separation trajectory of external devices based on a small sample size, which can accurately predict the separation trajectory of external devices when the amount of data is limited, and solve the problem that traditional methods are difficult and costly to obtain the separation trajectory of external devices.

[0005] Technical solution of the present invention: In order to achieve the above-mentioned purpose, according to the first aspect of the present invention, in order to achieve the above-mentioned purpose, the present invention adopts the following technical solution:

[0006] A method for predicting the trajectory of external storage release separation based on a small sample size is proposed. The specific steps include:

[0007] Step 1: Select N external attachments, and obtain the shape coordinates of each external attachment through experimental testing or numerical simulation methods, as well as the separation trajectory corresponding to Nt separation time steps and the aerodynamic history of the external attachment during the separation process; the separation trajectory is the coordinates of the center of gravity of the external attachment corresponding to Nt separation time steps; the aerodynamic history is the set of aerodynamic forces of the external attachment at each time step.

[0008] In a possible embodiment, N does not exceed 100. When N does not exceed 100, the method can achieve trajectory prediction with a small sample size.

[0009] In a possible embodiment, the specific process of obtaining the shape coordinates by the numerical simulation method includes:

[0010] The geometric model of the external attachment is imported into the meshing software, the surface and flow field of the external attachment are meshed, and then the coordinates of the mesh points on the surface of the external attachment are extracted as the external shape coordinates of the external attachment.

[0011] In a possible embodiment, the specific process of obtaining the separation trajectory and the aerodynamic history of the external attachment during the separation process by a numerical simulation method includes:

[0012] Computational fluid dynamics software is used to read the flow field grid of the external attachment; then the release and separation conditions of the external attachment are set, including Mach number, angle of attack, sideslip angle, altitude, etc., the aerodynamic force on the external attachment in the current time step is calculated, and the position of the external attachment at the next moment is updated according to the aerodynamic force on the external attachment; the process is repeated until the separation motion simulation of all Nt time steps is completed; the coordinates of the center of gravity position of the external attachment are read at each time step to obtain the separation trajectory; the aerodynamic force on the external attachment at each time step is read as the aerodynamic history of the external attachment during the separation process.

[0013] Step 2: Create an aerodynamic prediction network model, use the external attachment's shape coordinates and trajectory obtained in step 1 as input, and use the external attachment's aerodynamic history as output to train the aerodynamic prediction network model;

[0014] In a possible embodiment, the aerodynamic force prediction network model specifically includes:

[0015] The shape coordinate input layer, flattening layer, fully connected layer 1, and then connected to the input point 1 of the merging layer. The trajectory input layer is connected to the long short-term memory network layer, and then connected to the input point 2 of the merging layer. The output point of the merging layer is connected to the fully connected layer 2, fully connected layer 3, fully connected layer 4, reorganization layer and output layer in sequence.

[0016] The flattening layer reshapes the input data dimension into 1 dimension;

[0017] The mathematical expression of the fully connected layer is:

[0018]

[0019] Among them, x i is the input value of the i-th neuron; w ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron;

[0020] The output gate expression of the long short-term memory network layer is:

[0021] o t =σ(W o ·[h t-1 , x t ]+b o ) (2)

[0022] Among them, t is the output of the output gate at time step t; W o and b o are the weight matrix and bias of the output gate respectively; h t-1 is the hidden state output at time step t-1;

[0023] The reshaping layer reshapes the data dimension to be consistent with the output layer dimension;

[0024] Step 3: Create a trajectory prediction network model, use the external attachment's shape coordinates and the external attachment's aerodynamic history obtained in step 1 as input, and use the attachment's trajectory as output to train the trajectory prediction network model;

[0025] In a possible embodiment, the trajectory prediction network model specifically includes:

[0026] The shape coordinate input layer, flattening layer, fully connected layer 1, and then connected to the input point 1 of the merging layer. The aerodynamic process input layer is connected to the long short-term memory network layer, and then connected to the input point 2 of the merging layer. The output point of the merging layer is sequentially connected to the fully connected layer 2, fully connected layer 3, fully connected layer 4, reorganization layer and output layer.

[0027] The flattening layer reshapes the input data dimension into 1 dimension;

[0028] The mathematical expression of the fully connected layer is:

[0029]

[0030] Among them, x i is the input value of the i-th neuron; w ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron;

[0031] The output gate expression of the long short-term memory network layer is:

[0032] o t =σ(W o ·[h t-1 , x t ]+b o ) (4)

[0033] Among them, t is the output of the output gate at time step t; W o and b o are the weight matrix and bias of the output gate respectively; h t-1 is the hidden state output at time step t-1;

[0034] Step 4: Select an external attachment, initialize the trajectory history and aerodynamic history of the external attachment to 0, take the external attachment's shape coordinates and the external attachment's trajectory as input, and predict and update the external attachment's aerodynamic history through the trained aerodynamic prediction network model; take the external attachment's shape coordinates and the external attachment's aerodynamic history as input, and predict and update the external attachment's trajectory through the trained trajectory prediction network model; iterate this process, predict and update the aerodynamic history and trajectory in turn, until the preset number of iterations is reached; after the iteration is completed, the trajectory output by the trajectory prediction network model is the external attachment's release and separation trajectory.

[0035] Preferably, in step 2, the aerodynamic prediction network model requires input data sets and output data sets for training to achieve the prediction effect. The data obtained in step 1 are sorted into the input data sets and output data sets required for the training of the aerodynamic prediction network model. The input data set includes the external object shape coordinate matrix and the separation trajectory history matrix. Among them, the external object shape coordinate matrix is ​​an N×Np×n matrix, representing N external object samples, each sample is used to describe the shape with Np point clouds, and the coordinate dimension of each point cloud is n. For the two-dimensional case, n=2, and for the three-dimensional case, n=3. The separation trajectory matrix is ​​an N×Nt×m matrix, representing N external object samples, and the separation time of each sample is described by Nt discrete time steps, and the trajectory data dimension of each time step is m. For the two-dimensional case, m=3, and for the three-dimensional case, m=6. The output data set includes the external object aerodynamic history matrix, which is an N×Nt×m matrix, representing N external object samples, and the separation time of each sample is described by Nt discrete time steps, and the aerodynamic data dimension of each time step is m. For the two-dimensional case, m=3, and for the three-dimensional case, m=6.

[0036] Preferably, in step 2, the aerodynamic force prediction network model uses mean square error MSE as a loss function, and MSE is defined as:

[0037]

[0038] Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

[0039] Preferably, in step 2, the optimization training method of the aerodynamic force prediction network model is the Adam algorithm.

[0040] Preferably, in step 3, the trajectory prediction network model requires input data sets and output data sets for training to achieve the prediction effect. The data obtained in step 1 are sorted into the input data sets and output data sets required for the trajectory prediction network model training. The input data set includes the external object shape coordinate matrix and the aerodynamic history matrix. Among them, the external object shape coordinate matrix is ​​an N×Np×n matrix, representing N external object samples, each sample is used to describe the shape with Np point clouds, and the coordinate dimension of each point cloud is n. For the two-dimensional case, n=2, and for the three-dimensional case, n=3. The aerodynamic history matrix is ​​an N×Nt×m matrix, representing N external object samples, and the separation time of each sample is described by Nt discrete time steps, and the aerodynamic data dimension of each time step is m. For the two-dimensional case, m=3, and for the three-dimensional case, m=6. The output data set includes the external object trajectory matrix, which is an N×Nt×m matrix, representing N external object samples, and the separation time of each sample is described by Nt discrete time steps, and the trajectory data dimension of each time step is m. For the two-dimensional case, m=3, and for the three-dimensional case, m=6.

[0041] Preferably, in step 3, the trajectory prediction network model uses mean square error MSE as the loss function, and MSE is defined as:

[0042]

[0043] Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

[0044] Preferably, in step 3, the optimization training method of the trajectory prediction network model is the Adam algorithm.

[0045] Preferably, in step 4, the preset number of iterations is not less than 10 times.

[0046] The advantages of the present invention are: 1. The method of the present invention can achieve rapid and accurate prediction of the external store release separation trajectory with a small sample size. By establishing an external store aerodynamic prediction model and a trajectory prediction model, and adopting an iterative prediction method, the dependence on the external store release separation data set is effectively reduced, and the prediction accuracy and efficiency are improved. Specifically, in the experimental verification of the method of the present invention, in the case of 80 samples in the data set, the prediction results are almost consistent with the CFD calculation results. At the same time, the calculation speed is greatly improved compared to the CFD calculation, and the trajectory prediction is completed within 1 second.

[0047] 2. Compared with directly predicting the external attachment separation trajectory based on the external attachment shape coordinates, the method described in the present invention improves the accuracy of the trajectory prediction model in small sample conditions by establishing an external attachment aerodynamic prediction model and predicting the aerodynamic history of the external attachment as a feature supplement to the external attachment shape coordinates. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of an iterative prediction process of a method for predicting the trajectory of external attachment release and separation based on a small sample size according to the method described in this application;

[0049] Figure 2 The appearance of all external objects in the specific embodiment of the method described in this application;

[0050] Figure 3 The two-dimensional trajectory and posture prediction effect of the specific embodiment of the method described in this application;

[0051] In the figure, the trajectory prediction of four external storage samples is carried out. The horizontal and vertical coordinates and attitude angles of the center of gravity of the external storage are denoted as x, y, and Theta respectively. The CFD calculation results are represented by straight lines, and the predicted value data are represented by dots. Figure 4 In order to compare the two-dimensional trajectory and attitude prediction effects, trajectory prediction was performed on four external attachment samples in the figure. The horizontal and vertical coordinates of the center of gravity and the attitude angle of the external attachment are denoted as x, y, and Theta respectively. The CFD calculation results are represented by straight lines, and the predicted value data are represented by dots. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0053] The features and illustrative embodiments of various aspects of the present invention will be described in detail below. In the detailed description below, many specific details are proposed to provide a comprehensive understanding of the present invention. However, it is obvious to those skilled in the art that the present invention can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present invention by illustrating examples of the present invention. The present invention is by no means limited to any specific settings and methods proposed below, but covers any improvements, replacements and modifications of structures, methods, devices without departing from the spirit of the present invention. In the accompanying drawings and the following description, known structures and technologies are not shown to avoid unnecessary ambiguity in the present invention.

[0054] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other, and the embodiments can refer to and quote each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0055] Example

[0056] A method for predicting the optimal center of gravity position of an aircraft based on a deep neural network. The prediction process diagram is shown in Figure 1 As shown. The following steps are included:

[0057] Step 1: Select 80 external objects. The appearance of all external objects is as follows: Figure 2 As shown. The geometric model of the external attachment is imported into the meshing software, the surface and flow field of the external attachment are meshed, and then the coordinates of a total of 256 grid points on the surface of the external attachment are extracted as the external attachment's shape coordinates; the computational fluid dynamics software is used to read the flow field grid of the external attachment; then the external attachment's release and separation conditions are set, including Mach number, angle of attack, sideslip angle, altitude, etc., the aerodynamic force on the external attachment in the current time step is calculated, and the position of the external attachment at the next moment is updated according to the aerodynamic force on the external attachment; the process is repeated until all 26 time steps of separation motion simulation are completed; the horizontal and vertical coordinates of the center of gravity and the attitude angle of the external attachment are read at each time step to obtain the separation trajectory; the aerodynamic force on the external attachment at each time step is read as the aerodynamic history of the external attachment during the separation process;

[0058] Step 2: Create an aerodynamic force prediction network model, which specifically includes:

[0059] The shape coordinate input layer, flattening layer, fully connected layer 1, and then connected to the input point 1 of the merging layer. The trajectory input layer is connected to the long short-term memory network layer, and then connected to the input point 2 of the merging layer. The output point of the merging layer is connected to the fully connected layer 2, fully connected layer 3, fully connected layer 4, reorganization layer and output layer in sequence.

[0060] The flattening layer reshapes the input data dimension into 1 dimension;

[0061] The mathematical expression of the fully connected layer is:

[0062]

[0063] Among them, x i is the input value of the i-th neuron; w ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron;

[0064] The output gate expression of the long short-term memory network layer is:

[0065] o t =σ(W o ·[h t-1 , x t ]+b o ) (2)

[0066] Among them, t is the output of the output gate at time step t; W o and b o are the weight matrix and bias of the output gate respectively; h t-1 is the hidden state output at time step t-1;

[0067] The reshaping layer reshapes the data dimension to be consistent with the output layer dimension;

[0068] The data obtained in step 1 are organized into the input data set and output data set required for the training of the aerodynamic prediction network model. The input data set contains the external storage object shape coordinate matrix and the separation trajectory history matrix. Among them, the external storage object shape coordinate matrix is ​​a 80×256×2 matrix, representing 80 external storage object samples, each sample uses 256 point clouds to describe the shape, and the coordinate dimension of each point cloud is 2. The separation trajectory matrix is ​​a 80×26×3 matrix, representing 80 external storage object samples, the separation time of each sample is described by 26 discrete time steps, and the trajectory data dimension of each time step is 3. The output data set contains the external storage object aerodynamic history matrix, which is a 80×26×3 matrix, representing 80 external storage object samples, the separation time of each sample is described by 26 discrete time steps, and the aerodynamic data dimension of each time step is 3.

[0069] The aerodynamic prediction network model uses mean square error MSE as the loss function, and MSE is defined as:

[0070]

[0071] Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

[0072] The optimization training method of the aerodynamic prediction network model is the Adam algorithm.

[0073] Step 3: Create a trajectory prediction network model, which specifically includes:

[0074] The shape coordinate input layer, flattening layer, fully connected layer 1, and then connected to the input point 1 of the merging layer. The aerodynamic process input layer is connected to the long short-term memory network layer, and then connected to the input point 2 of the merging layer. The output point of the merging layer is sequentially connected to the fully connected layer 2, fully connected layer 3, fully connected layer 4, reorganization layer and output layer.

[0075] The flattening layer reshapes the input data dimension into 1 dimension;

[0076] The mathematical expression of the fully connected layer is:

[0077]

[0078] Among them, x i is the input value of the i-th neuron; w ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron;

[0079] The output gate expression of the long short-term memory network layer is:

[0080] o t =σ(W o ·[h t-1 , x t ]+b o ) (5)

[0081] Among them, t is the output of the output gate at time step t; W o and b o are the weight matrix and bias of the output gate respectively; h t-1 is the hidden state output at time step t-1;

[0082] The data obtained in step 1 is organized into the input data set and output data set required for trajectory prediction network model training. The input data set contains the external storage shape coordinate matrix and the aerodynamic history matrix. Among them, the external storage shape coordinate matrix is ​​a 80×256×2 matrix, representing 80 external storage samples, each sample uses 256 point clouds to describe the shape, and the coordinate dimension of each point cloud is 2. The aerodynamic history matrix is ​​a 80×26×3 matrix, representing 80 external storage samples, the separation time of each sample is described by 26 discrete time steps, and the aerodynamic data dimension of each time step is 3. The output data set contains the external storage trajectory matrix, which is a 80×26×3 matrix, representing 80 external storage samples, the separation time of each sample is described by 26 discrete time steps, and the trajectory data dimension of each time step is 3;

[0083] The trajectory prediction network model uses mean square error MSE as the loss function, and MSE is defined as:

[0084]

[0085] Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

[0086] The optimization training method of the trajectory prediction network model is the Adam algorithm.

[0087] Step 4: Select an external attachment, initialize the trajectory history and aerodynamic history of the external attachment to 0, take the external attachment's shape coordinates and the external attachment's trajectory as input, and predict and update the external attachment's aerodynamic history through the trained aerodynamic prediction network model; take the external attachment's shape coordinates and the external attachment's aerodynamic history as input, and predict and update the external attachment's trajectory through the trained trajectory prediction network model; iterate this process, predict and update the aerodynamic history and trajectory in turn, until the preset number of iterations of 50 times is reached; after the iteration is completed, the trajectory output by the trajectory prediction network model is the external attachment's release and separation trajectory.

[0088] Use Python to simulate the above steps, and select four test samples for prediction and comparison. Figure 3 The trajectory prediction results are shown in the figure. The trajectory prediction is performed on 4 external storage samples. The horizontal and vertical coordinates and attitude angles of the center of gravity of the external storage are recorded as x, y, and Theta respectively. The CFD calculation results are represented by straight lines, and the predicted value data are represented by dots. It can be seen that the predicted value of the trajectory proposed by the application is very close to the true value.

[0089] Comparative Example

[0090] The difference between the comparative example and the embodiment 1 is that the aerodynamic force prediction network model is not constructed, and the trajectory prediction model is directly used to perform separation trajectory prediction using the coordinates of the external attachment shape as input, which specifically includes the following steps:

[0091] Step 1: Select 80 external objects. The appearance of all external objects is as follows: Figure 2 As shown. The geometric model of the external attachment is imported into the meshing software, the surface and flow field of the external attachment are meshed, and then the coordinates of a total of 256 grid points on the surface of the external attachment are extracted as the external attachment's shape coordinates; the flow field grid of the external attachment is read using computational fluid dynamics software; then the external attachment's release and separation conditions are set, including Mach number, angle of attack, sideslip angle, altitude, etc., the aerodynamic force on the external attachment in the current time step is calculated, and the position of the external attachment at the next moment is updated according to the aerodynamic force on the external attachment; the process is repeated until all 26 time steps of separation motion simulation are completed; the horizontal and vertical coordinates of the center of gravity and the attitude angle of the external attachment are read at each time step to obtain the separation trajectory;

[0092] Step 2: Create a trajectory prediction network model, which specifically includes:

[0093] Shape coordinate input layer, flattening layer, fully connected layer 1, fully connected layer 2, fully connected layer 3, fully connected layer 4, reorganization layer and output layer;

[0094] The flattening layer reshapes the input data dimension into 1 dimension;

[0095] The mathematical expression of the fully connected layer is:

[0096]

[0097] Among them, x i is the input value of the i-th neuron; w ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron;

[0098] The output gate expression of the long short-term memory network layer is:

[0099] o t =σ(W o ·[h t-1 , x t ]+b o ) (2)

[0100] Among them, t is the output of the output gate at time step t; W o and b o are the weight matrix and bias of the output gate respectively; h t-1 is the hidden state output at time step t-1;

[0101] The data obtained in step 1 is organized into the input data set and output data set required for trajectory prediction network model training. The input data set contains the external hanger shape coordinate matrix. The external hanger shape coordinate matrix is ​​a 80×256×2 matrix, representing 80 external hanger samples. Each sample uses 256 point clouds to describe the shape, and the coordinate dimension of each point cloud is 2. The output data set contains the external hanger trajectory matrix, which is a 80×26×3 matrix, representing 80 external hanger samples. The separation time of each sample is described by 26 discrete time steps, and the trajectory data dimension of each time step is 3;

[0102] The trajectory prediction network model uses mean square error MSE as the loss function, and MSE is defined as:

[0103]

[0104] Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

[0105] The optimization training method of the trajectory prediction network model is the Adam algorithm.

[0106] Step 3: Select an external device, take the external device's shape coordinates as input, and predict the trajectory of the external device through the trained trajectory prediction network model. The trajectory output by the trajectory prediction network model is the release and separation trajectory of the external device.

[0107] Use Python to simulate the above steps, and select four test samples for prediction and comparison. Figure 4 The trajectory prediction results are shown in the figure. The trajectory prediction is performed on 4 external objects. The horizontal and vertical coordinates and attitude angles of the center of gravity of the external objects are recorded as x, y, and Theta respectively. The CFD calculation results are represented by straight lines, and the predicted value data are represented by dots. It can be seen that the predicted value of the trajectory by the comparison method has a large error with the true value. Compared with the comparison method, the method described in the present invention has higher prediction accuracy.

Claims

1. A method for predicting the trajectory of external attachment release and separation based on a small sample size, characterized in that: The specific steps include: Step 1: Select N external attachments, and obtain the shape coordinates of each external attachment through experimental testing or numerical simulation methods, as well as the separation trajectory corresponding to Nt separation time steps and the aerodynamic history of the external attachment during the separation process; the separation trajectory is the coordinates of the center of gravity of the external attachment corresponding to Nt separation time steps; the aerodynamic history is the set of aerodynamic forces of the external attachment at each time step; Step 2: Create an aerodynamic prediction network model, use the external attachment's shape coordinates and trajectory obtained in step 1 as input, and use the external attachment's aerodynamic history as output to train the aerodynamic prediction network model; Step 3: Create a trajectory prediction network model, use the external attachment's shape coordinates and the external attachment's aerodynamic history obtained in step 1 as input, and use the attachment's trajectory as output to train the trajectory prediction network model; Step 4: Select an external attachment, initialize the trajectory history and aerodynamic history of the external attachment to 0, take the external attachment's shape coordinates and the external attachment's trajectory as input, and predict and update the external attachment's aerodynamic history through the trained aerodynamic prediction network model; take the external attachment's shape coordinates and the external attachment's aerodynamic history as input, and predict and update the external attachment's trajectory through the trained trajectory prediction network model; iterate this process, predict and update the aerodynamic history and trajectory in turn, until the preset number of iterations is reached; after the iteration is completed, the trajectory output by the trajectory prediction network model is the external attachment's release and separation trajectory.

2. According to the method for predicting the trajectory of the external attachment release separation based on a small sample size according to claim 1, it is characterized in that: N does not exceed 100.

3. According to the method of claim 1, the method is characterized in that: The specific process of obtaining the shape coordinates by numerical simulation method includes: The geometric model of the external attachment is imported into the meshing software, the surface and flow field of the external attachment are meshed, and then the coordinates of the mesh points on the surface of the external attachment are extracted as the external shape coordinates of the external attachment.

4. According to the method of claim 1, the method is characterized in that: The specific process of obtaining the separation trajectory and the aerodynamic history of the external attachment during the separation process by numerical simulation methods includes: using computational fluid dynamics software to read the flow field grid of the external attachment; then setting the external attachment release and separation conditions, calculating the aerodynamic force on the external attachment at the current time step, and updating the position of the external attachment at the next moment according to the aerodynamic force on the external attachment; repeating this process until the separation motion simulation of all Nt time steps is completed; reading the coordinates of the center of gravity position of the external attachment at each time step to obtain the separation trajectory; reading the aerodynamic force on the external attachment at each time step as the aerodynamic history of the external attachment during the separation process.

5. According to the method for predicting the trajectory of external attachment release separation based on a small sample size in claim 1, it is characterized in that: The aerodynamic force prediction network model specifically includes: The shape coordinate input layer, flattening layer, fully connected layer 1, and then connected to the input point 1 of the merging layer. The trajectory input layer is connected to the long short-term memory network layer, and then connected to the input point 2 of the merging layer. The output point of the merging layer is connected to the fully connected layer 2, fully connected layer 3, fully connected layer 4, reorganization layer and output layer in sequence. The flattening layer reshapes the input data dimension into 1 dimension; The mathematical expression of the fully connected layer is: Among them, x i is the input value of the i-th neuron; w ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron; The output gate expression of the long short-term memory network layer is: the t =σ(W o [h t-1 ,x t ]+b o ) (2) Among them, t is the output of the output gate at time step t; W o and b o are the weight matrix and bias of the output gate respectively; h t-1 is the hidden state output at time step t-1; the reshaping layer reshapes the data dimension to be consistent with the output layer dimension.

6. The method for predicting the trajectory of external attachment release and separation based on a small sample size according to claim 1 is characterized in that: The trajectory prediction network model specifically includes: The shape coordinate input layer, flattening layer, fully connected layer 1, and then connected to the input point 1 of the merging layer. The aerodynamic process input layer is connected to the long short-term memory network layer, and then connected to the input point 2 of the merging layer. The output point of the merging layer is sequentially connected to the fully connected layer 2, fully connected layer 3, fully connected layer 4, reorganization layer and output layer. The flattening layer reshapes the input data dimension into 1 dimension; The mathematical expression of the fully connected layer is: Among them, x i is the input value of the i-th neuron; W ji is the weight connecting the i-th input neuron and the j-th output neuron; b j is the bias of the jth output neuron; The output gate expression of the long short-term memory network layer is: the t =σ(W o ·[h t-1 ,x t ]+b o ) (4) Among them, t is the output of the output gate at time step t; W o and b o are the weight matrix and bias of the output gate respectively; h t-1 is the hidden state output at time step t-1.

7. The method for predicting the trajectory of external attachment release and separation based on a small sample size according to claim 1 is characterized in that: In step 2, the data obtained in step 1 are organized into input data sets and output data sets required for aerodynamic prediction network model training. The input data set includes an external attachment shape coordinate matrix and a separation trajectory history matrix, wherein the external attachment shape coordinate matrix is ​​an N×Np×n matrix, representing N external attachment samples, each sample uses Np point clouds to describe the shape, and the coordinate dimension of each point cloud is n. For the two-dimensional case, n=2, and for the three-dimensional case, n=3; the separation trajectory matrix is ​​an N×Nt×m matrix, representing N external attachment samples, and the separation time of each sample is described by Nt discrete time steps, and the trajectory data dimension of each time step is m; for the two-dimensional case, m=3, and for the three-dimensional case, m=6; the output data set includes an external attachment aerodynamic history matrix, which is an N×Nt×m matrix, representing N external attachment samples, and the separation time of each sample is described by Nt discrete time steps, and the aerodynamic data dimension of each time step is m; for the two-dimensional case, m=3, and for the three-dimensional case, m=6.

8. The method for predicting the trajectory of external attachment release and separation based on a small sample size according to claim 1 is characterized in that: In step 2, the aerodynamic prediction network model uses the mean square error MSE as the loss function, and MSE is defined as: Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

9. The method for predicting the trajectory of external attachment release and separation based on a small sample size according to claim 1 is characterized in that: In step 3, the data obtained in step 1 are organized into input data sets and output data sets required for trajectory prediction network model training; the input data set includes an external attachment shape coordinate matrix and an aerodynamic history matrix; wherein the external attachment shape coordinate matrix is ​​an N×Np×n matrix, representing N external attachment samples, each sample uses Np point clouds to describe the shape, and the coordinate dimension of each point cloud is n, for the two-dimensional case, n=2, for the three-dimensional case, n=3; the aerodynamic history matrix is ​​an N×Nt×m matrix, representing N external attachment samples, the separation time of each sample is described by Nt discrete time steps, and the aerodynamic data dimension of each time step is m; for the two-dimensional case, m=3, for the three-dimensional case, m=6; the output data set includes an external attachment trajectory matrix, which is an N×Nt×m matrix, representing N external attachment samples, the separation time of each sample is described by Nt discrete time steps, and the trajectory data dimension of each time step is m; for the two-dimensional case, m=3, for the three-dimensional case, m=6.

10. The method for predicting the trajectory of external attachment release and separation based on a small sample size according to claim 1 is characterized in that: In step 3, the trajectory prediction network model uses the mean square error MSE as the loss function, and MSE is defined as: Where n is the number of data samples; y i is the true value of the i-th data sample; is the predicted value of the ith data sample.

Images