Target flight trajectory prediction compensation method based on residual network

Abstract

The invention discloses a target flight path prediction compensation method based on a residual network, and the method mainly comprises the steps: designing a volume information filtering method according to the current and historical flight paths of a target, and obtaining a target flight path reference prediction value; designing a prediction compensation network structure based on a residual module, and building a target real flight trajectory, a Kalman filtering prediction trajectory and a Kalman filtering prediction error data set; selecting an optimizer suitable for the network, adjusting network training hyper-parameters, including initialization parameters of the network optimizer, batch size, learning rate and the like, completing convergence training of the network in a training set, and outputting a prediction compensation value for a target flight trajectory by the network; and performing comprehensive calculation on the volume information filtering algorithm reference prediction value and the residual network prediction compensation value to obtain a final trajectory prediction result. The key point of the method is the compensation network design, training and deployment reasoning based on the residual module, and the prediction precision of the flight path of the high-dynamic target can be effectively improved.

Description

Technical Field

[0001] The present invention designs a target flight trajectory prediction method based on artificial intelligence, in particular a prediction compensation method based on a residual network on the basis of filtering prediction, which belongs to the engineering application field of artificial intelligence technology. Background Art

[0002] The task of predicting the flight trajectory of an enemy target is to reasonably calculate and infer the possible location of the target aircraft in the future based on the motion characteristics and historical motion trajectory of the target aircraft. In an informationized battlefield environment, accurate prediction results of the target flight trajectory can provide reliable data support for tactical decision-making. Excellent trajectory prediction capabilities can enable us to form an OODA (Observation, Orientation, Decision and Action) cycle more quickly and establish advantages in complex and ever-changing air combat environments. Therefore, it is urgent to enhance the ability to predict the flight trajectory of highly dynamic targets in complex and highly dynamic battlefield environments.

[0003] Current research results can be roughly divided into two categories. One is a prediction method based on knowledge-driven, such as the paper "Hypersonic Aircraft Track Estimation Based on Learnable EKF" published by Zheng Tianyu in the Journal of Harbin Institute of Technology. It established a near-space hypersonic target maneuvering model and proposed a learnable Kalman filter trajectory prediction method based on the maneuvering model. However, this method needs to combine the motion model of the predicted target to optimally estimate the motion state of the target at the future moment, but the dynamic equation of the target is difficult to obtain in the actual scene. The other is a prediction method based on data-driven, represented by deep learning algorithms, which establishes a prediction model of the target trajectory through a large amount of historical flight data of the target. For example, the patent published by Guo Jifeng (CN112348223A) discloses a missile flight trajectory prediction method based on deep learning, and the patent published by He Ping (CN114637334A) proposes an intelligent UAV flight trajectory prediction method based on the Informer network. This method no longer uses numerical integration methods that require a large number of repeated calculations, and at the same time gets rid of the dependence on the target motion model. However, these methods all have the disadvantages of large consumption of computing resources, difficulty in extracting trajectory data features, and difficulty in network training convergence. Therefore, the current flight trajectory prediction methods for highly dynamic targets are still insufficient. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of data-driven intelligent methods in the target flight trajectory prediction task, such as large consumption of computing resources, difficulty in extracting trajectory data features, and difficulty in network convergence. A target flight trajectory prediction compensation method based on residual network is proposed to improve the prediction accuracy of the target flight trajectory.

[0005] Compared with the prior art, the present invention has at least the following beneficial effects:

[0006] 1) Aiming at targets with different maneuvering motion characteristics in air combat environments, a prediction error compensation model is established by combining the characteristics of the existing filtering prediction algorithm to improve the prediction accuracy of the existing trajectory prediction algorithm;

[0007] 2) An intelligent network framework with a residual short-circuit mechanism is proposed, which transforms the network learning objective from learning feature mapping to learning residual, reduces the difficulty of network training convergence, improves the stability of network training, and solves the problem of network degradation.

[0008] 3) A network structure with convolutional layers and long short-term memory layers was designed to extract features of the target flight trajectory from the spatial and temporal dimensions, enriching the data feature dimensions and improving the accuracy of residual learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The target flight trajectory prediction and compensation method based on residual network of the present invention is given by the following embodiments and drawings

[0010] Figure 1 A flow chart of a target flight trajectory prediction and compensation method based on a residual network of the present invention;

[0011] Figure 2 This is a diagram of a flight trajectory prediction and compensation network structure based on a residual module according to an embodiment of the present invention;

[0012] Figure 3 This is a target flight trajectory prediction result diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0014] The present invention discloses a target flight trajectory prediction and compensation method based on a residual network, which mainly includes: designing a volume information filtering method to obtain a target flight trajectory benchmark prediction value according to the current and historical flight trajectory of the target; designing a prediction compensation network structure based on a residual module, building a target real flight trajectory, a Kalman filter prediction trajectory, and a Kalman filter prediction error data set; selecting a suitable optimizer for the network, adjusting the network training hyperparameters, including network optimizer initialization parameters, batch size, learning rate, etc., completing the convergence training of the network in the training set, and the network outputs the prediction compensation value for the target flight trajectory; performing a comprehensive calculation on the volume information filtering algorithm benchmark prediction value and the residual network prediction compensation value to obtain the final trajectory prediction result. The present invention focuses on the compensation network design, training and deployment reasoning based on the residual module, which can effectively improve the prediction accuracy of the flight trajectory of a high-dynamic target.

[0015] Example 1

[0016] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The implementation steps of the present invention are as follows:

[0017] Step 1: Establish a target trajectory dataset to simulate the target's motion in real scenarios. Target motion models are established for escaping targets (S-shaped maneuvers), decoy missiles (parabolic), and interceptor missiles (proportional guidance), and target flight trajectory datasets are generated based on the three types of maneuvers.

[0018] Step 2: Preprocess the target flight trajectory data. First, use a sliding window to segment the trajectory data. Move the window with the same step size until the entire trajectory is traversed. Then normalize the obtained segmented data. The specific normalization formula is as follows:

[0019]

[0020] At the same time, the data of the time length that needs to be predicted is taken as the corresponding output, and the same segmentation and normalization processing is completed, and the preprocessed input and output trajectories are saved correspondingly.

[0021] Step 3: Design a local filter. Combining the advantages of volumetric Kalman filtering and information filtering, a volumetric information filtering algorithm is proposed. Based on the volumetric information filtering algorithm, a local filter is designed to complete the filtering prediction of the target flight trajectory and obtain the prediction reference value. The specific design of the volumetric information filtering algorithm is as follows.

[0022] 1) Create m (m = 2n) volume points

[0023]

[0024] In the formula

[0025]

[0026] where e s represents an n-dimensional unit vector, e s The sth element of is 1, and the rest are 0.

[0027] 2) Time update

[0028] The one-step prediction state vector can be calculated as follows

[0029]

[0030] Then the one-step prediction information error covariance matrix and its corresponding one-step prediction information state vector The calculation is as follows

[0031]

[0032] 3) Observation update

[0033] Calculate another set of volume points

[0034]

[0035] Then the one-step prediction observation vector can be calculated according to the following formula

[0036]

[0037] The measurement error covariance matrix is ​​obtained from this and the one-step-forecast cross-covariance matrix

[0038]

[0039] 4) Filter update prediction

[0040] Defining and calculating information contribution vectors and information contribution matrix as follows

[0041]

[0042] The state estimation results can be obtained:

[0043]

[0044] Then the estimated value of the system state at time k can be obtained according to the following formula:

[0045]

[0046] Step 4: Design a target flight trajectory prediction compensation network with a residual structure. Design a one-dimensional convolutional layer (1DCNN) and a long short-term memory layer (LSTM) to complete the feature extraction of the target flight trajectory. At the same time, based on the residual structure, a short-circuit mechanism is added to the network to change the learning goal of the network from learning feature mapping to learning residuals, which solves the problem of network degradation and improves the convergence speed and stability of the network. The complete structure of the compensation network is shown in the attached figure. Figure 2 ;

[0047] Step 5: Complete the training of the target flight trajectory prediction compensation network, set and adjust the network's hyperparameters, including network initialization parameters, batch size, learning rate, etc. The specific network settings here are as follows. The network contains two one-dimensional convolutional layers, one maximum pooling layer, and three LSTM layers. The number of neurons in each layer is set to 128→64→8, and the training data input batchsize is set to 256. The network as a whole uses the adaptive moment estimation (Adam) optimizer to optimize the network training. The initial momentum is set to 0.9 and the learning rate is set to 0.0005. By adjusting the network's hyperparameters, the network's training convergence is accelerated, the network's fitting accuracy is improved, and the network's inference deployment is completed at the same time to obtain the target flight trajectory prediction compensation value;

[0048] Step 6: The predicted reference value obtained by comprehensive volume information filtering and the predicted compensation value obtained by compensation network reasoning are used to calculate the final predicted value of the target flight trajectory.

Claims

1. A target flight trajectory prediction and compensation method based on residual network, characterized in that: The following steps are involved: Step 1: Establish a target trajectory dataset, simulate the target's motion in real scenarios, establish target motion models for escaping targets, decoy missiles, and interceptor missiles, and generate target flight trajectory datasets based on the three maneuver types; Step 2: Preprocess the target flight trajectory data; Step 3: Design a local filter. Combining the advantages of volumetric Kalman filtering and information filtering, a volumetric information filtering algorithm is proposed. Based on the volumetric information filtering algorithm, a local filter is designed to complete the filtering prediction of the target flight trajectory and obtain the prediction reference value. Step 4: Design a target flight trajectory prediction compensation network with a residual structure, and change the network's learning goal from learning feature mapping to learning residuals; Step 5: Complete the training of the target flight trajectory prediction compensation network and obtain the target flight trajectory prediction compensation value; Step 6: The predicted reference value obtained by comprehensive volume information filtering and the predicted compensation value obtained by compensation network reasoning are used to calculate the final predicted value of the target flight trajectory.

2. A target flight trajectory prediction and compensation method based on residual network as claimed in claim 1, characterized in that: The target motion model in step 1 is escaping target-S-type maneuver, decoy missile-parabola, interceptor missile-proportional guidance.

3. A target flight trajectory prediction and compensation method based on residual network as claimed in claim 1, characterized in that: In step 2, the trajectory data is first segmented using a sliding window, the window is moved with the same step size until the entire trajectory is traversed, and the obtained segmented data is normalized.

4. A target flight trajectory prediction and compensation method based on residual network as claimed in claim 3, characterized in that: The normalization processing formula is as follows: Among them, μ i is the mean of the data, δ i is the standard deviation of the data; at the same time, the data of the time length that needs to be predicted is taken as the corresponding output, and the same segmentation and normalization processing is completed, and the preprocessed input and output trajectories are saved correspondingly.

5. A target flight trajectory prediction and compensation method based on residual network as claimed in claim 4, characterized in that: The volume information filtering algorithm in step 3 is specifically as follows: 1) Create m (m = 2n) volume points In the formula where e s represents an n-dimensional unit vector, e s The sth element of is 1, and the rest are 0; 2) Time update The one-step prediction state vector can be calculated as follows Then the one-step prediction information error covariance matrix and its corresponding one-step prediction information state vector The calculation is as follows: 3) Observation update Calculate another set of volume points Then the one-step prediction observation vector can be calculated according to the following formula The measurement error covariance matrix is ​​obtained from this and the one-step-forecast cross-covariance matrix 4) Filter update prediction Defining and calculating information contribution vectors and information contribution matrix as follows The state estimation results can be obtained: Then the estimated value of the system state at time k can be obtained according to the following formula:

6. A target flight trajectory prediction and compensation method based on residual network as claimed in claim 5, characterized in that: In step 4, a one-dimensional convolution layer and a long short-term memory layer are designed to complete the feature extraction of the target flight trajectory. At the same time, a short-circuit mechanism is added to the network based on the residual structure to convert the learning goal of the network from learning feature mapping to learning residual.

7. A target flight trajectory prediction and compensation method based on residual network as claimed in claim 6, characterized in that: In step 5, the hyperparameters of the network are set and adjusted, including network initialization parameters, batch size, and learning rate.

8. A target flight trajectory prediction and compensation method based on residual network as claimed in claim 7, characterized in that: The network is set up as follows: the network includes two one-dimensional convolutional layers, one maximum pooling layer, and three LSTM layers. The number of neurons in each layer is set to 128→64→8, and the training data input batch size is set to 256. The network as a whole uses an adaptive momentum estimation optimizer to optimize the network training. The initial momentum is set to 0.9 and the learning rate is set to 0.0005.

9. A target flight trajectory prediction and compensation method based on residual network as claimed in claim 8, characterized in that: By adjusting the network's hyperparameters, the network's training convergence is accelerated, the network's fitting accuracy is improved, and at the same time, the network's reasoning deployment is completed to obtain the target flight trajectory prediction compensation value.

Images