An aerial formation prediction method and system
By combining LSTM and SVM models, a formation grid coding method was developed to solve the challenges of feature extraction and template construction in aerial formation prediction, enabling intelligent prediction of formations and improving the accuracy and efficiency of formation prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- AIR FORCE EARLY WARNING ACADEMY
- Filing Date
- 2022-07-27
- Publication Date
- 2026-07-03
AI Technical Summary
In existing technologies, aerial formation prediction algorithms face challenges in feature extraction and template construction. Traditional methods are complex and computationally expensive, making it difficult to meet the needs of formation trajectory prediction.
By combining LSTM and SVM models, and through formation grid encoding and transfer learning, we can quickly extract the physical spatial features of the formation and predict the formation at the next moment.
It enables intelligent prediction of aerial formations, improves the accuracy and efficiency of formation prediction, and provides a useful reference for intelligent perception of air combat situation.
Smart Images
Figure CN115270626B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of situational awareness, and more specifically, relates to an aerial formation prediction method and system. Background Technology
[0002] Aerial formation prediction, also known as aerial formation forecasting, is crucial because aerial formations are often related to the enemy's aerial actions, objectives, and missions. Predicting formations allows for early assessment of the air combat situation and provides an advantage in air combat. However, due to the real-time nature of formation changes, it is difficult for humans to predict formations accurately and promptly based solely on experience.
[0003] Formation prediction is based on formation recognition, and traditional computer-aided formation recognition methods are template-based. In template-based methods, formation feature extraction is crucial for recognition. Existing feature extraction methods based on Hough transform and others are complex and difficult to adjust parameters, severely impacting formation recognition performance. Furthermore, template construction is highly subjective and involves considering numerous factors. All of these limitations restrict the application of traditional template-based recognition methods.
[0004] In formation prediction, predicting the trajectories of individual aircraft within a formation is crucial. Traditional trajectory prediction methods are primarily based on physical modeling, generally employing two approaches: numerical and analytical. Numerical methods incur significant computational costs, while analytical methods require the construction of complex models; neither can adequately meet the demands of formation trajectory prediction. Summary of the Invention
[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an aerial formation prediction method and system, which aims to solve the problems of difficult feature extraction and template construction in existing formation prediction algorithms.
[0006] To achieve the above objectives, the present invention provides an aerial formation prediction method, comprising the following steps:
[0007] S1: Using the extreme values of latitude and longitude of each target coordinate in the current formation as the boundary, and taking the movement direction of the lead aircraft in the formation as the formation movement direction, select the coordinates of the upper left and lower right vertices of the formation area to construct the formation area coordinates and complete the construction of the formation area;
[0008] S2: Divide the formation area into grids, and encode each grid according to whether a target exists in each grid to form a formation grid code; wherein, each grid contains at most one target;
[0009] S3: Concatenate the array grid codes from multiple consecutive time steps into feature vectors and input them into the trained LSTM model to predict the feature vector corresponding to the array grid code at the next time step.
[0010] S4: Input the feature vector corresponding to the formation grid code at the next moment into the SVM model to predict the formation at the next moment;
[0011] S5: Use the next time step's formation grid code and formation as the current time step's formation grid code and formation, respectively. Then input the feature vectors corresponding to the formation grid codes of multiple consecutive time steps into the LSTM model. Slide forward with a sliding window of length N and stride of 1 to obtain the feature vectors corresponding to the formation grid codes of the next time step.
[0012] S6: Input the feature vector corresponding to the formation grid code at the next moment into the SVM model to obtain the formation at the next moment; go back to S5 until no formation recognition stops.
[0013] More preferably, the formation grid code R a×b In the grid, 1 indicates that a target has fallen into the grid, and 0 indicates that no target has fallen into the grid; the elements in the formation grid code are:
[0014]
[0015] Where a is the matrix dimension of the formation's direction of movement; b is the matrix dimension perpendicular to the formation's direction of movement.
[0016] More preferably, the formation includes: wedge formation, column formation, echelon formation, horizontal formation, or serpentine formation.
[0017] More preferably, based on the measured formation grid code, LSTM and SVM models are trained using transfer learning methods to output the predicted formation; wherein, the kernel function and parameters of the SVM model are obtained by network search and cross-validation.
[0018] On the other hand, the present invention provides an aerial formation prediction system, comprising:
[0019] The formation area construction module is used to construct the formation area coordinates by taking the extreme values of the latitude and longitude of each target coordinate in the formation at the current moment as the edge, taking the movement direction of the lead aircraft in the formation as the formation movement direction, selecting the coordinates of the upper left and lower right vertices of the formation area, and completing the construction of the formation area.
[0020] The encoding module is used to divide the formation area into grids and encode the grids according to whether there is a target in each grid, forming a formation grid code; wherein, each grid contains at most one target.
[0021] The LSTM driver module is used to concatenate the array grid codes from multiple consecutive time steps into feature vectors and input them into the trained LSTM model. The model then slides forward with a sliding window of length N and stride of 1 to predict the feature vector corresponding to the array grid code at the next time step.
[0022] The SVM driver module is used to input the feature vector corresponding to the formation grid code at the next time step into the SVM model to predict the formation at the next time step.
[0023] The update module is used to use the formation grid code and formation shape of the next time step as the formation grid code and formation shape of the current time step, respectively.
[0024] More preferably, the formation grid code R a×b In the grid, 1 indicates that a target has fallen into the grid, and 0 indicates that no target has fallen into the grid; the elements in the formation grid code are:
[0025]
[0026] Where a is the matrix dimension of the formation's direction of movement; b is the matrix dimension perpendicular to the formation's direction of movement.
[0027] More preferably, the formation includes: wedge formation, column formation, echelon formation, horizontal formation, or serpentine formation.
[0028] More preferably, based on the measured formation grid code, LSTM and SVM models are trained using transfer learning methods to output the predicted formation; wherein, the kernel function and parameters of the SVM model are obtained by network search and cross-validation.
[0029] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art:
[0030] This invention proposes an aerial formation prediction method and system. The invention provides an aerial formation grid encoding method to extract the physical spatial features of aerial formations. Formation grid codes from multiple consecutive time points are concatenated into feature vectors and input into a trained LSTM model to predict the feature vector corresponding to the formation grid code at the next time point. The feature vector corresponding to the formation grid code at the next time point is then input into an SVM model to predict the formation configuration at the next time point. Transfer learning is used to train the LSTM and SVM models, outputting the predicted formation configuration. Specific example experiments verify the effectiveness of the aerial formation prediction method provided by this invention. It achieves intelligent aerial formation prediction, directly deriving the formation configuration at the next time point, providing a valuable reference for intelligent situational awareness in air combat. Attached Figure Description
[0031] Figure 1(a) is an aerial formation situation diagram provided in an embodiment of the present invention;
[0032] Figure 1(b) is a schematic diagram of the queuing grid code provided in an embodiment of the present invention;
[0033] Figure 2This is a schematic diagram of the formation prediction model (LSTM model combined with SVM model) provided in an embodiment of the present invention;
[0034] Figure 3 This is a schematic diagram of a network-based deep transfer learning method provided in an embodiment of the present invention;
[0035] Figure 4(a) is a schematic diagram of the wedge formation provided in an embodiment of the present invention;
[0036] Figure 4(b) is a schematic diagram of the column formation provided in an embodiment of the present invention;
[0037] Figure 4(c) is a schematic diagram of the echelon formation provided in an embodiment of the present invention;
[0038] Figure 4(d) is a schematic diagram of the horizontal formation provided in an embodiment of the present invention;
[0039] Figure 5 This is a schematic diagram of the formation area coordinates provided in an embodiment of the present invention;
[0040] Figure 6 This is a schematic diagram of the queuing grid provided in an embodiment of the present invention;
[0041] Figure 7(a) is a schematic diagram of the formation flight trajectory provided in an embodiment of the present invention;
[0042] Figure 7(b) is a diagram showing the formation prediction effect provided by an embodiment of the present invention.
[0043] Figure 8(a) is a schematic diagram of the SVM model recognition accuracy when the number of samples in each class is 50, as provided in the embodiment of the present invention;
[0044] Figure 8(b) is a schematic diagram of the SVM model recognition accuracy when the number of samples in each class is 1000, as provided in the embodiment of the present invention. Detailed Implementation
[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0046] This invention provides an aerial formation prediction method and system, wherein the training process of the model is as follows: First, the physical space features of the formation are extracted quickly; then, based on the aerial formation grid coding, an intelligent formation prediction model is designed based on an LSTM network; simultaneously, an intelligent recognition model based on the SVM algorithm is also used; finally, the model is trained using transfer learning.
[0047] In practical applications, aerial formation prediction methods include the following steps:
[0048] S1: Using the extreme values of latitude and longitude of each target coordinate in the current formation as the boundary, and taking the movement direction of the lead aircraft in the formation as the formation movement direction, select the coordinates of the upper left and lower right vertices of the formation area to construct the formation area coordinates and complete the construction of the formation area;
[0049] S2: Divide the formation area into grids and encode each grid according to whether a target exists in it, forming a binary formation grid code; wherein, each grid contains at most one target;
[0050] S3: Concatenate the array grid codes from multiple consecutive time steps into feature vectors and input them into the trained LSTM model to predict the feature vector corresponding to the array grid code at the next time step.
[0051] S4: Input the feature vector corresponding to the formation grid code at the next moment into the SVM model to predict the formation at the next moment;
[0052] S5: Use the next time step's formation grid code and formation as the current time step's formation grid code and formation, respectively. Input the feature vectors corresponding to the formation grid codes of multiple consecutive time steps into the LSTM model, and slide forward with a sliding window of length N and stride of 1 to obtain the feature vector corresponding to the formation grid code of the next time step.
[0053] S6: Input the feature vector corresponding to the formation grid code at the next moment into the SVM model to obtain the formation at the next moment; go back to S5 until no formation recognition stops.
[0054] More preferably, the formation grid code R a×b In the grid, 1 indicates that a target has fallen into the grid, and 0 indicates that no target has fallen into the grid; the elements in the formation grid code are:
[0055]
[0056] Where a is the matrix dimension of the formation's direction of movement; b is the matrix dimension perpendicular to the formation's direction of movement.
[0057] More preferably, the formation includes wedge formation, column formation, echelon formation, line formation, or serpentine formation;
[0058] More preferably, based on the measured formation grid code, LSTM and SVM are trained using transfer learning methods to output the formation pattern; wherein, the SVM kernel function and its parameters are obtained by network search and cross-validation.
[0059] A more detailed explanation follows:
[0060] On the one hand, this embodiment provides an aerial formation prediction method, including the following steps:
[0061] Step 1: Building an aerial formation recognition model;
[0062] A model for aerial formation recognition is constructed. Formation recognition is a classification problem. The formation structure features are extracted through encoding and then classified by a classifier SVM.
[0063] (1.1) Raster coding
[0064] Formation recognition is based on formation feature extraction. Traditional methods exist for extracting formation features, but these methods suffer from problems such as complex design, numerous human factors, and high computational cost. To address this issue, this invention proposes a grid-coded formation feature extraction method, which facilitates the rapid and convenient extraction of spatial features of formations.
[0065] First, the direction of movement of the lead aircraft in the formation is taken as the formation's direction of movement. Then, a square formation area is formed by taking the extreme values of the latitude and longitude of each target in the formation at the same moment as the edges. Along the formation's direction of movement, the coordinates of the upper left and lower right vertices of this area are taken to construct the formation area coordinates (P). l P u Q l Q u After obtaining the formation area, the area is divided into grids and binary encoded to form a uniform binary matrix, which is the formation grid code. In the grid code, 1 indicates that the target falls into this grid and 0 indicates that there is no target in this grid, as shown in Figure 1(a) and Figure 1(b).
[0066] Figure 1(a) shows the aerial formation situation, and Figure 1(b) shows the raster code corresponding to Figure 1(a). It can be seen that the raster code simplifies the airspace situation and maps it to the coding space; for the raster code R... a×b Where a is the matrix dimension of the formation's direction of movement; b is the matrix dimension perpendicular to the formation's direction of movement; the matrix elements can be represented as:
[0067]
[0068] Aerial formations can be categorized into dense, sparse, and dispersed formations based on the density of targets. The main criterion for distinguishing between these formations is their density. Correspondingly, grid codes can be categorized into three types: dense formation grid codes, sparse formation grid codes, and dispersed formation grid codes. The type of grid code used during the encoding process can be determined by setting a threshold.
[0069] (1.2) Model Construction
[0070] SVM is an effective statistical learning method that can automatically find support vectors with good discriminative ability for classification. The classifier constructed from this can maximize the inter-class margin and effectively solve problems such as small sample size, nonlinearity, high dimensionality, and local minima. In formation recognition practice, there are problems such as limited sample data and imperfectly linearly separable sample attributes; the SVM algorithm is well-suited for such application scenarios.
[0071] For the queuing dataset (x) i y i ), x i ∈R m , i = 0, 1, ..., n, x i ,y i Let x represent the sample feature vector and class representation, respectively, where m and n represent the sample feature dimension and the number of samples, respectively; the sample feature vector x i This refers to the array grid code, where m is the encoding length. When the target features are not perfectly linearly separable, slack variables can be introduced to construct a soft margin and find its optimal classification surface. The mathematical form of this problem is:
[0072]
[0073] Where w is the direction vector; b is the bias; ε i The slack variable corresponds to each sample point; the constant C is the penalty parameter, which controls the degree of penalty for misclassified samples; after constructing the Lagrangian function, its dual problem becomes:
[0074]
[0075] The optimal Lagrange operator solution λ is obtained by the SMO algorithm. * And obtain the optimal classification hyperplane w * x+b * If = 0, then the optimal classification function becomes:
[0076]
[0077] When the target features are linearly inseparable, a kernel function can be introduced to map the features to a high-dimensional space and then find its optimal classification surface.
[0078] Step 2: Formation prediction model construction;
[0079] The formation prediction model takes formation data from multiple consecutive time steps as input to predict the formation at the next time step; here, the formation prediction model is built based on LSTM.
[0080] LSTM is a typical example of a recurrent neural network (RNN). By introducing the concepts of gating mechanisms and unit states, it avoids the problems of gradient explosion and long-term dependencies, and can better "learn" long-term patterns. It is widely used in processing time-series data. The LSTM model is a single-layer structure with N (number of observations) LSTM neurons. The model structure is as follows: Figure 2 As shown;
[0081] In an LSTM network, the input to an LSTM unit includes the current input x. t The output h of the previous unit t-1 and unit state C t-1 The current input consists of three parts; in the LSTM unit, a forget gate f is introduced. t Input gate i t and output gate o t To regulate the flow of output information, the forget gate determines whether to save or discard information. The expression is:
[0082] f t =sigmoid(W f ·[h t-1 x t ]+b f (5)
[0083] Among them, W f and b f The weights and biases of the forget gate;
[0084] The input gate is used to update the state C of this unit. t The expression is:
[0085] C t =C t-1 ⊙f t +i t ⊙tanh(W c ·[h t-1 x t ]+b c (6)
[0086] Among them, i t =sigmoid(W i ·[h t-1 x t ]+b i W i W c and b i b c These are the weights and biases associated with the input gate, respectively.
[0087] The output gate determines the output h of this unit. t The relevant expression is:
[0088] h t =O t ⊙tanh(C t (7)
[0089] Among them, O t =sigmoid(W o ·[h t-1 x t ]+b o )w o and b o C represents the weights and biases of the output gate; t with h t As input to the next LSTM unit;
[0090] For a formation consisting of P targets, assuming a total of N detections are performed (i.e., the sliding window length is N), the formation encoding feature vector from the k-th detection is input into the k-th LSTM unit. After prediction by the Long Short-Term Network (LSTM), the predicted formation encoding feature vector at time N+1 is obtained. The predicted formation encoding feature vector at time N+1 is then input into the SVM recognition model to obtain the formation type at time N+1. In the prediction of the next time step, a sliding window of length N and stride of 1 is used to move forward, and so on to obtain all subsequent prediction results.
[0091] Step 3: Training the formation recognition model;
[0092] For machine learning algorithms, a large amount of experimental data is fundamental, but it is difficult to obtain a large amount of experimental data in practice.
[0093] To address these issues, this invention employs a network-based deep transfer learning method for training LSTM and SVM models. Network-based deep transfer learning refers to reusing a pre-trained portion of a network in the source domain, including its network structure and connection parameters, and transferring it to a deep neural network used in the target domain. A schematic diagram of network-based deep transfer learning is shown below. Figure 3 As shown;
[0094] First, a deep learning network model is trained in the source domain using a large training dataset. Then, a portion of the pre-trained network in the source domain is transferred to the target domain, making it part of a new deep transfer learning network model. Finally, after training and adjustment, the network parameters are updated, and the deep transfer learning model is trained. The deep transfer learning network can be viewed as two parts: the front layer can be regarded as a feature extractor, and the back layer can be regarded as a classifier.
[0095] This embodiment uses simulator flight data to construct the source domain dataset; based on operational regulations, multiple people are organized to conduct formation flight through a flight simulator system to extract dense formation data in the air and establish a dataset; the target domain dataset consists of measured data and is used for fine-tuning the model;
[0096] When predicting real-time aerial formations, the data is first rasterized; then the coded data is input into the trained prediction model to obtain the formation.
[0097] On the other hand, the present invention provides an aerial formation prediction system, comprising:
[0098] The formation area construction module is used to construct the formation area coordinates by taking the extreme values of the latitude and longitude of each target coordinate in the formation at the current moment as the edge, taking the movement direction of the lead aircraft in the formation as the formation movement direction, selecting the coordinates of the upper left and lower right vertices of the formation area, and completing the construction of the formation area.
[0099] The encoding module is used to divide the formation area into grids and encode the grids according to whether there is a target in each grid, forming a formation grid code; wherein, each grid contains at most one target.
[0100] The LSTM driver module is used to concatenate the array grid codes from multiple consecutive time steps into feature vectors and input them into the trained LSTM model. The model then slides forward with a sliding window of length N and stride of 1 to predict the feature vector corresponding to the array grid code at the next time step.
[0101] The SVM driver module is used to input the feature vector corresponding to the formation grid code at the next time step into the SVM model to predict the formation at the next time step.
[0102] The update module is used to use the formation grid code and formation shape of the next time step as the formation grid code and formation shape of the current time step, respectively.
[0103] More preferably, the formation grid code R a×b In the grid, 1 indicates that a target has fallen into the grid, and 0 indicates that no target has fallen into the grid; the elements in the formation grid code are:
[0104]
[0105] Where a is the matrix dimension of the formation's direction of movement; b is the matrix dimension perpendicular to the formation's direction of movement.
[0106] More preferably, the formation includes: wedge formation, column formation, echelon formation, horizontal formation, or serpentine formation.
[0107] More preferably, based on the measured formation grid code, LSTM and SVM models are trained using transfer learning methods to output the predicted formation; wherein, the kernel function and parameters of the SVM model are obtained by network search and cross-validation.
[0108] Example
[0109] 1. Source Domain Data Generation
[0110] This embodiment is based on the operational regulations, organizing multiple people to conduct formation flying through a flight simulation system, extracting dense formation data in the air and establishing a dataset;
[0111] According to their shape, dense aerial formations can be divided into wedge formations, column formations, echelon formations, line formations, and serpentine formations. After multiple simulated flights, 4,000 sets of data of dense formations of 4 types (wedge formations, column formations, echelon formations, and line formations) were extracted, as shown in Figures 4(a) to (d). The dataset was divided into training set and test set in an 8:2 ratio.
[0112] 2. Raster coding
[0113] The formation's direction of motion is taken from the direction of motion of the lead aircraft. Along this direction, the coordinates of the top-left and bottom-right vertices of the region are used to construct the formation area coordinates. Figure 5 As shown;
[0114] In this embodiment, a four-aircraft formation was observed. At time t1, the extreme values of the latitude and longitude of each target in the formation were 125.01°E, 125.04°E, 28.53°N, and 28.54°N. Therefore, the coordinates of this formation area can be represented as (125.01°E, 125.04°E, 28.53°N, 28.54°N). After obtaining the formation area, the area was divided into a grid with a grid size of 5×6. The grid was then binary encoded to form a uniformly sized binary matrix, i.e., the formation grid code is as follows: Figure 6 As shown;
[0115] When inputting a recognition model built on SVM, the binary matrix form of the arrayed grid code can be converted into a one-dimensional vector.
[0116] 3. Model building and training
[0117] All models used in this invention are implemented using Python programming language. The LSTM neural network model is built using the PyTorch deep learning framework, and the SVM model is built using the Sklearn model.
[0118] The selection of the SVM kernel function and its parameters is crucial. Through grid search and cross-validation, the "Linear" kernel function is selected, with a penalty coefficient C of 0.2.
[0119] The LSTM network model parameters were randomly initialized, and the network model was trained using the Adam optimization method with a learning rate set to 10. -3 The batch size was set to 64, and the number of training rounds was set to 300.
[0120] 4. Prediction Results
[0121] Flight data for a specific formation was extracted using a flight simulation system, and formation recognition and testing were performed throughout the entire process. The formation consisted of four identical aircraft, and its flight trajectory is shown in Figure 7(a). A total of 85 observations were conducted from the start to the end of data extraction. Formation recognition was performed for each observation. During prediction, a sliding window of 8 was used, meaning prediction began from the 9th observation and continued until the formation flight ended. The recognition model parameters were those obtained from training with 1000 samples per class. The identification and prediction results are shown in Figure 7(b). In Figure 7(a), the entire track is divided into five segments based on the change in the formation's direction of motion: AB segment, where the formation flies in a horizontal formation (25 observations); BC segment, where the formation turns left (9 observations); CD segment, where the formation adjusts to a vertical formation (22 observations); DE segment, where the formation turns right (7 observations); EF segment, where the formation returns to a horizontal formation (22 observations). As can be seen in Figure 7(b), the AB track... In the first segment, the formation prediction made one error, predicting a wedge formation instead of a line formation. In the second segment, the formation turned left and failed to maintain a line formation; the identification model identified it as a trapezoidal formation, resulting in two prediction errors. In the third segment, the formation adjusted to a column formation, and the prediction matched the identification result. In the fourth segment, the formation turned right, and the prediction made one error. In the fifth segment, the formation returned to a line formation, and the prediction result was correct. A total of 77 predictions were made throughout the entire process, with 73 of them matching the identification result, achieving a prediction accuracy of 95%.
[0122] 5. Recognition performance
[0123] Recognition is the foundation of prediction, and the model's recognition performance directly determines the prediction result. To intuitively depict the performance of the recognition model, and considering the limited amount of real formation data, which can lead to small sample sizes, the model's recognition performance is highly sensitive to the sample size. Therefore, we plotted confusion matrices and calculated the model's recognition accuracy when the number of samples in each class was 50 and 1000, respectively. The experimental results are shown in Figures 8(a) and 8(b), where 0, 1, 2, and 3 represent wedge, trapezoid, horizontal, and vertical formations, respectively. The higher the value at the diagonal position of the confusion matrix, the better the model's recognition performance.
[0124] As shown in Figure 8(a), when the number of samples in each class is 50, the recognition model has a misclassification rate of 6%, 5%, 6%, and 4% for wedge, trapezoid, horizontal, and vertical formations, respectively, with a model recognition accuracy of 94.8%. As shown in Figure 8(b), when the number of samples in each class is 1000, the recognition model has a misclassification rate of 5%, 4%, 5%, and 4% for wedge, trapezoid, horizontal, and vertical formations, respectively, with a model recognition accuracy of 95.5%. The comparison shows that increasing the data sample size has limited effect on improving the model's recognition performance, demonstrating the excellent recognition capability of the SVM-based recognition model under small sample conditions.
[0125] This invention proposes an aerial formation prediction method and system. The invention provides an aerial formation grid encoding method to extract the physical spatial features of aerial formations. Formation grid codes from multiple consecutive time points are concatenated into feature vectors and input into a trained LSTM model to predict the feature vector corresponding to the formation grid code at the next time point. The feature vector corresponding to the formation grid code at the next time point is then input into an SVM model to predict the formation configuration at the next time point. Transfer learning is used to train the LSTM and SVM models, outputting the predicted formation configuration. Specific example experiments verify the effectiveness of the aerial formation prediction method provided by this invention. It achieves intelligent aerial formation prediction, directly deriving the formation configuration at the next time point, providing a valuable reference for intelligent situational awareness in air combat.
[0126] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting aerial formations, characterized in that, Includes the following steps: S1: Using the extreme values of latitude and longitude of each target coordinate in the current formation as the boundary, and taking the movement direction of the lead aircraft in the formation as the formation movement direction, select the coordinates of the upper left and lower right vertices of the formation area to construct the formation area coordinates and complete the construction of the formation area; S2: Divide the formation area into grids, and encode each grid according to whether a target exists in each grid to form a formation grid code; wherein, each grid contains at most one target; S3: Concatenate the array grid codes from multiple consecutive time steps into feature vectors and input them into the trained LSTM model to predict the feature vector corresponding to the array grid code at the next time step. S4: Input the feature vector corresponding to the formation grid code at the next moment into the SVM model to predict the formation at the next moment; S5: Use the next time step's formation grid code and formation as the current time step's formation grid code and formation, respectively. Then concatenate the formation grid codes of multiple consecutive time steps into feature vectors and input them into the LSTM model. Slide forward with a sliding window of length N and stride of 1 to obtain the feature vector corresponding to the formation grid code of the next time step. S6: Input the feature vector corresponding to the formation grid code at the next moment into the SVM model to obtain the formation at the next moment; go back to S5 until no formation recognition stops.
2. The aerial formation prediction method according to claim 1, characterized in that, Grouping grid code In the grid, 1 indicates that a target has fallen into the grid, and 0 indicates that no target has fallen into the grid; the elements in the formation grid code are: in, a The matrix dimension represents the direction of formation movement; b The matrix dimension is perpendicular to the direction of formation movement.
3. The aerial formation prediction method according to claim 1 or 2, characterized in that, Formations include: wedge formation, column formation, echelon formation, line formation, or serpentine formation.
4. The aerial formation prediction method according to claim 3, characterized in that, Based on the measured formation grid code, LSTM and SVM models are trained using transfer learning to output the predicted formation. The kernel function and parameters of the SVM model are obtained through network search and cross-validation.
5. An aerial formation prediction system, characterized in that, include: The formation area construction module is used to construct the formation area coordinates by taking the extreme values of the latitude and longitude of each target coordinate in the formation at the current moment as the edge, taking the movement direction of the lead aircraft in the formation as the formation movement direction, selecting the coordinates of the upper left and lower right vertices of the formation area, and completing the construction of the formation area. The encoding module is used to divide the formation area into grids and encode the grids according to whether there is a target in each grid, forming a formation grid code; wherein, each grid contains at most one target. The LSTM driver module is used to concatenate the array grid codes from multiple consecutive time steps into feature vectors and input them into the trained LSTM model. The model then slides forward with a sliding window of length N and stride of 1 to predict the feature vector corresponding to the array grid code at the next time step. The SVM driver module is used to input the feature vector corresponding to the formation grid code at the next time step into the SVM model to predict the formation at the next time step. The update module is used to use the formation grid code and formation shape of the next time step as the formation grid code and formation shape of the current time step, respectively.
6. The aerial formation prediction system according to claim 5, characterized in that, Grouping grid code In the grid, 1 indicates that a target has fallen into the grid, and 0 indicates that no target has fallen into the grid; the elements in the formation grid code are: in, a The matrix dimension represents the direction of formation movement; b The matrix dimension is perpendicular to the direction of formation movement.
7. The aerial formation prediction system according to claim 5 or 6, characterized in that, Formations include: wedge formation, column formation, echelon formation, line formation, or serpentine formation.
8. The aerial formation prediction system according to claim 7, characterized in that, Based on the measured formation grid code, LSTM and SVM models are trained using transfer learning to output the predicted formation. The kernel function and parameters of the SVM model are obtained through network search and cross-validation.