CN113393032A

CN113393032A - Flight path cycle prediction method based on resampling

Info

Publication number: CN113393032A
Application number: CN202110658739.1A
Authority: CN
Inventors: 刘向丽; 宋仪雯; 柯励; 李赞; 王志国; 李学楠
Original assignee: Xidian University; Xian Jiaotong University
Current assignee: Xidian University; Xian Jiaotong University
Priority date: 2021-06-15
Filing date: 2021-06-15
Publication date: 2021-09-14
Anticipated expiration: 2041-06-15
Also published as: CN113393032B

Abstract

The invention discloses a resampling-based track cycle prediction method, which mainly solves the problems that in the track prediction in the prior art, the prediction time is short and the prediction error is overlarge due to the change of the future motion state of a target. The implementation scheme is as follows: simulating a historical track and a future track of the maneuvering target; sequentially carrying out preprocessing of filtering, resampling and normalization on the historical track data of the target; constructing a neural network model consisting of a Bi-LSTM layer, a Dropout layer, a Dense layer and an activation layer, and training the neural network model by utilizing preprocessed track data; generating part of historical track data by using a circulation strategy, and calculating the historical track data by using trained neural network model parameters; and carrying out smooth filtering on the calculation result to obtain the final predicted flight path. The method has the advantages of small prediction error and long prediction time, can still obtain more accurate predicted flight path when the motion state of the target changes in the future, and can be used for target tracking.

Description

Flight path cycle prediction method based on resampling

Technical Field

The invention belongs to the technical field of communication, and particularly relates to a track circulation prediction method which can be used for target tracking.

Background

The target track prediction technology is used for accurately predicting the future track state information of a target and is one of key technologies in the field of target tracking.

With the trend of complicated aviation flying environment, the influence of uncertain factors such as system error, severe weather and abnormal performance of the sensor can cause the problem that the sensor can not continuously detect target track information and the flight safety is influenced. Therefore, a certain flight path prediction capability is required, more complete data information is provided for follow-up target tracking, and flight safety is ensured.

At present, flight path prediction algorithms are mainly divided into a dynamic model based on flight performance parameters, an optimal estimation model based on parameters and a machine learning method based on historical data.

In the aspect of a dynamic model, a flight path prediction method based on a large circular route and an equiangular route is provided in the application analysis of the paper flight path prediction method in the course of flight, and information such as the circular route and the equiangular route is introduced into the model. The method provides an aerodynamic-based track prediction method in a thesis aircraft runway sliding track prediction method, and carries out stress analysis on a target to establish a sliding dynamics model. Porretta in the paper Performance Evaluation of a Novel 4D Trajectory Prediction Model for circuit Aircraft proposed an Aircraft Performance Model that takes into account wind speed, lateral braking force of the Aircraft and speed estimates. However, model parameters such as weather forecast, scene control intention, flight plan information and the like required by these methods are difficult to obtain in actual application, so that in the case of loss of these model parameters, the target cannot be accurately modeled, which may result in inaccurate prediction results.

In terms of optimal parameter estimation, the most classical algorithm is target state estimation based on kalman filtering to predict the trajectory of the aircraft. The method is characterized in that the method provides an interactive multi-model combined unscented Kalman filtering algorithm in the airport surface moving target tracking based on IMM algorithm in the thesis, and the algorithm models the moving process of the airplane and carries out the prediction calculation of the track. The hybrid system theory-based flight path prediction algorithm is provided in the conflict-free 4D flight path prediction based on the hybrid system theory in the thesis, a parameter evolution model of an aircraft dynamics model is constructed according to the motion characteristics of an aircraft in different flight sections, a state transfer model is constructed to model the switching between different flight sections, and the flight path prediction of the aircraft is completed by adjusting corresponding parameters. However, the algorithm has low execution efficiency, and when the motion state of the target is not known, the target cannot be accurately modeled, so that the predicted track error is large. In the aspect of machine learning models, Malyong provides a precise track prediction method based on data mining in the research of a four-dimensional track precise prediction method based on data mining in a thesis. Such algorithms predict more accurately than classical prediction methods, but still have the problem of short prediction time.

In conclusion, the existing target track prediction technologies have the defects of large error, single prediction model and short prediction time, so that the accuracy of the predicted track is poor, and the flight safety is influenced.

Disclosure of Invention

The invention aims to provide a cycle prediction method based on resampling to reduce the prediction error of the maneuvering target track, increase the prediction time and improve the track prediction accuracy.

The technical scheme of the invention is that target historical track data is preprocessed; constructing and training a neural network model; generating part of historical track data by using a circulation strategy, and calculating the historical track data by using trained network model parameters; and carrying out smooth filtering on the calculation result to obtain the final predicted flight path.

According to the above thought, the flight path cycle prediction method based on resampling is characterized by comprising the following steps:

(1) filtering and normalizing historical track data of the maneuvering target;

(2) setting the resampling period T' as 10 times of the system sampling period T, sampling the normalized flight path data into 10 resampling flight paths as a network data set, and determining the input sample dimension and the label sample dimension of the network data set;

(3) constructing a neural network sequentially consisting of a Bi-directional long-short time memory unit Bi-LSTM layer, a Dropout layer, a Dense layer and an activation layer, wherein the four layers are of structures;

(4) setting the maximum iteration number as N and the batch size batch _ size, sending the network data set into the constructed network, performing iterative training on the parameters of the network by using a batch gradient descent method, and obtaining a trained network model when the iteration number reaches N;

(5) and predicting the flight path of a future period of time:

(5a) processing the 10 resampled flight path data by adopting a circulation strategy to generate 10 batches of data for predicting future flight paths;

(5b) calculating the 10 batches of data processed by the circulation strategy in sequence by calling trained neural network parameters to obtain 10 predicted flight paths, and combining the 10 flight paths into a predicted flight path according to the sequence of the original time slot before resampling;

(6) and carrying out smooth filtering processing on the predicted track by adopting a smooth method to obtain a final prediction result.

Compared with the prior art, the invention has the following advantages:

1) the invention uses Bi-LSTM layer of the neural network to bidirectionally read the historical track of the target and learn the change characteristics of the target, and when the motion state of the target changes in the future, the future track can still be predicted more accurately.

2) The invention uses the resampling strategy to resample the historical flight path in different time slots, so that the data set can reflect the change information of the historical flight path more uniformly, the prediction error of the neural network predicted flight path is reduced, and the prediction accuracy is further improved.

3) According to the method, a plurality of batches of data are generated by adopting a cyclic strategy for the resampled flight path data, and are sequentially predicted, compared with the traditional single prediction method, the method utilizes the data information of the known flight path to a greater extent, can obtain the predicted flight path with longer predicted time, provides a more accurate flight strategy for the target, and ensures the flight safety of the target.

Drawings

FIG. 1 is a flow chart of an implementation of the present invention;

fig. 2 is a diagram of a neural network model architecture used in the present invention.

Detailed Description

Embodiments and effects of the present invention will be described in further detail below with reference to the accompanying drawings.

Referring to fig. 1, the flight path cycle prediction method based on resampling of the present invention includes the following steps:

step 1, generating a track data set.

Setting the system duration to be 3000s, the system sampling period T to be 1s, the process evolution noise variance to be 0.01, the target x-axis initial speed to be 200m/s, the y-axis initial speed to be 200m/s, the z-axis initial speed to be 0m/s, the x-axis initial coordinate to be 15, the y-axis initial coordinate to be 15 and the z-axis initial coordinate to be 15000 m;

the target makes periodic motion with a period of 600s, and the maneuvering state parameters in a single period are shown in a table 1:

TABLE 1 partial time horizon target maneuver information

Table 1 shows the maneuvering state change of the target in the first period (0-600 s), the uniform motion of 300s, the left turning motion of 150s and the right turning motion of 150s in a single period. The target entire trajectory took five cycles, lasting 3000 s. And the front 2400s of the target track is set as a known historical track, and the rear 600s is set as an unknown future track.

Observing a target through a radar sensor observation platform, setting observation sampling frequency to be 1s, setting radar delay to be 1s, adding observation noise based on radar angle and radius, and setting the noise size to be 0.001 degrees and 100 m;

and (3) fitting the motion state of the target by using an interactive multi-model algorithm to generate the front 2400s known track data and the rear 600s unknown track data of the target.

And 2, performing Kalman filtering processing on the known track information in front of the target 2400 s.

(2.1) calculating the state prediction value of the target next time k

Sum error covariance P_k|k-1：

Where k is the discrete time period, F is the target state transition matrix, T is the matrix transpose,

is a current time state value, P_k-1|k-1The error covariance at the current moment is Q, and a process noise covariance matrix is Q;

(2.2) calculating track State update values

Sum error covariance update value P_k|k：

P_k|k＝[I-[P_k|k-1H^T(HP_k|k-1H^T+R)^-1]H]P_k|k-1，

Wherein H is an observation matrix of the flight path, R is a measurement noise covariance matrix, and Z_kThe data are measured by the radar at the moment k, and I is a unit array;

and (2.3) repeating the steps (2.1) and (2.2) for 2400 times to obtain the filtered flight path.

Step 3, normalization processing is carried out on the filtered flight path to obtain normalized flight path data X_s：

Wherein x is_iIs the ith time slot data of the filtered track.

Step 4, normalizing the track data X_sSampling into 10 resampling tracks.

The specific implementation of this step is as follows:

1, selecting the 1 st time slot of the original flight path as a starting point by using the 1 st resampling flight path, and resampling with a resampling interval T' of 10 s;

selecting the 2 nd time slot of the original flight path as a starting point by using the 2 nd resampling flight path, and resampling with a resampling interval T' of 10 s;

in the same way, the 10 th resampling track selects the 10 th time slot of the original track as the starting point, and resampling with the resampling interval T' of 10s is carried out;

each resampled flight path contains 300 sampling points, and the corresponding relation between the time slots of 10 resampled flight paths and the time slots of the original flight path is obtained, as shown in table 2:

TABLE 2 corresponding table of resampled flight path and time slot of original flight path

And 5, determining the network input sample dimension and the label sample dimension.

The specific implementation of this step is as follows:

for the determination of the network input sample dimension: firstly, determining the length W of a sliding window of an input sample as 120 and the number of features as 1; sequentially performing sliding segmentation on the 10 resampled flight path data according to W, and splicing segmentation results to obtain a network input sample with sample batch of 610 and dimension of 610 × 120 × 1;

for the determination of label exemplar dimensions: setting the sample batch and the feature number of the label samples to be the same as those of the input samples, and setting the time step number of each batch of samples to be 60 to obtain the label samples with the dimension of 610 × 60 × 1;

the resampled track time slots corresponding to the input samples and the tag samples are shown in table 3:

TABLE 3 input sample/tag sample and track time slot correspondence table

And 6, constructing a neural network model.

Referring to fig. 2, the neural network model constructed in this step is sequentially composed of four layers of structures, from top to bottom, including a Bi-LSTM layer, a Dropout layer, a sense layer, and an activation layer, and each layer has the following functions and parameters:

the Bi-LSTM layer is used for extracting the change characteristics of the historical track data set, and the hidden node number units of the Bi-LSTM layer is 200;

a Dropout layer for preventing overfitting of the network in the training process, wherein the drop rate Dropout _ ratio is 0.2;

the Dense layer is used for fitting a label sample Y _ train during network training, and the hidden node number units of the Dense layer is 60;

and the activation layer is used for enhancing the adaptability of the network model to nonlinear data, and the activation function of the activation layer is a linear activation function.

And 7, training a neural network model.

The existing method for training the neural network comprises a batch gradient descent method, a random gradient descent method and a mini gradient descent method, wherein the step adopts but is not limited to the batch gradient descent method, and the method is realized as follows:

(7.1) setting the batch size of data batch _ size to be 64, dividing the track data set into a plurality of small batches of data according to the batch size of the data, and sequentially sending the small batches of data into a neural network for single training;

(7.2) setting a network optimization algorithm as an adaptive moment estimation algorithm Adam, and optimizing network parameters by calculating and correcting the first moment and the second moment of each round of training gradient;

and (7.3) setting the maximum iteration number of the network to be 100, and repeating the steps (7.1) and (7.2) to reach the maximum iteration number to obtain the trained network model.

And 8, generating 10 batches of data for predicting future tracks.

The specific implementation of this step is as follows:

(8.1) duration T predicted according to flight path requirement_preResampling period T' and system sampling period T, calculating predicted batch N for 10 batches of data:

N＝T_pre/T/T'，

wherein, T in this example_preThe time is 600s, T is 1s, T' is 10s, and the predicted batch N is 60 according to a formula;

(8.2) selecting 60 test data batches for 10 resampling tracks by adopting a cyclic strategy to generate 10 data batches:

selecting track data of 62 th to 181 th time slots of the 1 st resampled track for the 1 st batch of test data of the 1 st batch of data;

selecting track data of 63 st to 182 th time slots of the 1 st re-sampling track from the 2 nd batch of test data of the 1 st batch of data;

in the same way, for the 60 th batch of test data of the 1 st batch of data, selecting the track data of the 121 st to 240 th time slot of the 1 st re-sampling track;

selecting flight path data of 62 th to 181 th time slots of the 2 nd resample flight path from the 1 st batch of test data of the 2 nd batch of data;

selecting track data of 63 rd to 182 th time slots of the 2 nd resampling track for the 2 nd batch of test data of the 2 nd batch of data;

and in the same way, for the 10 th batch of 60 th test data, selecting the track data of 121 th to 240 th time slots of the 10 th resampled track.

The 60 batches of test data correspond to 10 track time slots after resampling, as shown in table 4:

TABLE 4 track time slot mapping table after network input resampling

And 9, sequentially calculating the 10 batches of data processed by the circulation strategy by calling the trained neural network parameters to obtain 10 predicted flight paths.

(9.1) calculating in sequence 60 test data of the 1 st data obtained in (8.2):

calling the trained network parameters to calculate the 1 st batch of test data of the 1 st batch of data to obtain a flight path of 182-241 time slots;

calling the trained network parameters to calculate the 2 nd batch of test data of the 1 st batch of data to obtain a flight path of 183-242 time slots;

by analogy, the trained network parameters are called to calculate the 60 th batch of test data of the 1 st batch of data to obtain the flight path of 241-300 time slots,

the above-mentioned 60 prediction results obtained in total correspond to the resampled track time slot, as shown in table 5:

table 560 sets prediction results and resampled track time slot corresponding table

(9.2) sequentially selecting the last time slot point of the 60 batches of prediction results, splicing the last time slot point into a prediction track containing 60 time slot points, namely selecting the last time slot point 241 of the prediction results in the 1 st prediction batch, selecting the last time slot point 242 of the prediction results in the 2 nd prediction batch, and so on, selecting the last time slot point 300 of the prediction results in the 60 th prediction batch, and sequentially splicing the data into track data containing 60 time slot points, namely the prediction track of the 1 st data.

(9.3) predicting the data of the 2 nd to 9 th batches by adopting the methods of (9.1) and (9.2) to obtain the predicted flight path of the data of the 2 nd to 9 th batches,

and (9.4) combining the 10 predicted tracks obtained from (9.1) to (9.3) into one predicted track according to the original time slot sequence before resampling.

And step 10, smoothing filtering the predicted track by adopting a smooth method to obtain a final prediction result.

The formula of the smoothing filter is as follows:

yy(n)＝(y(1)+y(2)+y(3)+...+y(n))/n，

wherein y (n) represents a numerical value before the smoothing of the nth element, yy (n) represents a numerical value after the smoothing of the nth element, and the predicted track after the smoothing is the final prediction result.

The effects of the present invention can be further illustrated by the following simulation experiments.

Firstly, simulation conditions:

the neural network model takes a Keras framework of Python3.6 as a simulation platform;

the flight path generation part and the Kalman filtering part take MATLAB 2018a as a simulation platform;

secondly, simulation content:

under the above conditions, the present invention and the two existing track prediction algorithms are adopted to respectively predict the target future track for 600s for a long time, and the prediction results are obtained, as shown in table 6:

TABLE 6 comparison of different track prediction algorithm results

As can be seen from Table 6, compared with the existing algorithm, the flight path cycle prediction method based on resampling provided by the invention has the advantages that the error is lower when the flight path is predicted, the prediction time is longer, a more accurate flight strategy can be provided for the target, and the flight safety of the target is ensured.

The above description is only a specific example of the present invention and should not be construed as limiting the invention in any way, and it will be apparent to those skilled in the art that various modifications and variations in form and detail can be made without departing from the principle and structural concept of the present invention after understanding the present invention, but such modifications and variations are still within the scope of the appended claims.

Claims

1. A flight path cycle prediction method based on resampling is characterized by comprising the following steps:

(1) filtering and normalizing historical track data of the maneuvering target;

(5) and predicting the flight path of a future period of time:

(5b) calculating the 10 batches of data processed by the circulation strategy in sequence by calling trained neural network parameters to obtain 10 predicted flight paths, and combining the predicted flight paths into a predicted flight path according to the time slot sequence of the original flight path before resampling;

2. The method according to claim 1, wherein the filtering process for the flight path data in (1) is to perform state prediction and state update for the historical flight path data by using a kalman filtering algorithm to eliminate the measurement noise of the system, and is implemented as follows:

(1a) calculating a state prediction value of a target next time k

Sum error covariance P_k|k-1：

Wherein k is the discrete time period, F is the target state transition matrix, and Q is the system process noise.

(1b) Calculating track state update values

With error coordinationVariance update value P_k|k：

P_k|k＝[I-G_kH_k]P_k|k-1

H is an observation matrix of the track, R (k) is a noise covariance matrix, T is a matrix transposition, and Z is radar measurement data;

(1c) and (4) performing iterative calculation on the (1a) and the (1b), wherein the iteration times are the time slot points of the historical track, and obtaining the filtered target track.

3. The method of claim 1, wherein the track data is normalized in (1) by the following equation:

wherein x is_iFor the ith time-slot data of the filtered track, X_scaledIs normalized data.

4. The method of claim 1, wherein (2) the normalized track data is sampled into 10 resampled tracks as follows:

1, the 1 st resampling track selects the 1 st time slot of the original track as a starting point, and resampling with a sampling interval of T' is carried out;

the 2 nd resampling track selects the 2 nd time slot of the original track as a starting point, and resampling with a sampling interval of T' is carried out;

in the same way, the 10 th resampling track selects the 10 th time slot of the original track as the starting point, and resampling with the sampling interval of T' is carried out, so that the original track is changed into the 10 resampling tracks.

5. The method of claim 1, wherein the network input sample dimension and the label sample dimension are determined in (2) by:

for the determination of the network input sample dimension, firstly, the length W of a sliding window of the input sample is determined to be 120, the number of features is 1, then, 10 pieces of flight path data after resampling are sequentially subjected to sliding segmentation according to W, segmentation results are spliced, and the network input sample with the sample batch of 610 and the dimension of 610 × 120 × 1 is obtained.

For the determination of the dimension of the label sample, the sample batch and the feature number of the label sample are set to be the same as those of the input sample, and the time step number of each batch of samples is set to be 60, so that the label sample with the dimension of 610 × 60 × 1 can be obtained.

6. The method of claim 1, wherein the neural network constructed in (3) has the following functions and parameters:

7. The method of claim 1, wherein the parameters of the neural network are iteratively trained by using a small batch gradient descent method in (4), and the following steps are implemented:

(4a) setting the batch size of data to 64, dividing a track data set into a plurality of small batches of data according to the batch size of the data, and sequentially sending the small batches of data into a neural network for single training;

(4b) setting a network optimization algorithm as an adaptive moment estimation algorithm Adam, and optimizing network parameters by calculating and correcting a first moment and a second moment of each round of training gradient;

(4c) and (4) setting the maximum iteration number of the network as 100, and repeating the step (4a) and the step (4b) for 100 times to obtain a trained network model.

8. The method of claim 1, wherein the generating of the 10 batches of data for predicting future tracks in (5a) is performed as follows:

(5a1) duration T predicted according to flight path requirement_preResampling period T' and system sampling period T, calculating predicted batch N for each batch in 10 batches of data:

N＝T_pre/T/T'；

(5a2) and selecting N batches of network inputs by adopting a cyclic strategy for 10 resampling tracks to generate 10 batches of data:

in the same way, selecting the track data of the 121 th to 240 th time slot of the 1 st resampling track for the Nth batch of test data of the 1 st batch of data;

in the same way, for the test data of the Nth batch of the 10 th batch, the track data of the 121 th to 240 th time slot of the 10 th resampling track is selected.

9. The method of claim 1, wherein 10 batches of data are sequentially calculated in (5b) by calling the trained neural network parameters, and the following steps are performed:

(5b1) and (5) calling the trained network parameters to sequentially calculate the N batches of test data of the 1 st batch of data obtained in the step (5a) to obtain N batches of prediction results.

(5b2) Sequentially selecting the last time slot point of the N batches of prediction results, and splicing the last time slot points into a prediction track containing N time slot points, namely the prediction track of the 1 st batch of data;

(5b3) and (5b1) and (5b2) are adopted to predict the data of the 2 nd to 9 th batches, so that predicted tracks of the data of the 2 nd to 9 th batches are obtained, and 10 predicted tracks are combined into a predicted track according to the original time slot sequence before resampling.

10. The method according to claim 1, wherein in (6), smooth filtering processing is performed on the predicted track by using a smooth method, and the formula is as follows:

yy(n)＝(y(1)+y(2)+y(3)+...+y(n))/n