CN113393032B - Track cycle prediction method based on resampling - Google Patents

Abstract

The invention discloses a track circulation prediction method based on resampling, which mainly solves the problems of short prediction time length and overlarge prediction error caused by the change of a target in a future motion state in track prediction in the prior art. The implementation scheme is as follows: simulating a historical track and a future track of the maneuvering target; the target historical track data is subjected to pretreatment of filtering, resampling and normalization in sequence; 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 the preprocessed track data; generating partial historical track data by using a circulation strategy, and calculating the partial historical track data by using trained neural network model parameters; and carrying out smooth filtering on the calculation result to obtain a final predicted track. The method has the advantages of smaller prediction error and longer prediction time, can still obtain more accurate prediction tracks when the motion state of the target changes in the future, and can be used for target tracking.

Description

Track cycle prediction method based on resampling

Technical field

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

Background technique

Target trajectory prediction technology is to accurately predict the target's future trajectory status information and is one of the key technologies in the field of target tracking.

As the aviation flight environment becomes more complex, due to the influence of uncertain factors such as system errors, bad weather, and abnormal performance of the sensor itself, the sensor will be unable to continue to detect target track information, affecting flight safety. Therefore, it is necessary to have certain trajectory prediction capabilities to provide more complete data information for subsequent target tracking and ensure flight safety.

At present, trajectory prediction algorithms are mainly divided into dynamic models based on flight performance parameters, parameter optimal estimation models and machine learning methods based on historical data.

In terms of dynamic models, Fu Qiang proposed a trajectory prediction method based on great circle routes and equiangular routes in his paper Analysis of the Application of Track Prediction Methods in Route Flights, and introduced information such as circular routes and equiangular routes into the model. . In his paper, aircraft runway taxiing trajectory prediction method, Liao Chaowei proposed a trajectory prediction method based on aerodynamics. This method analyzes the force of the target and establishes a taxiing dynamics model. In the paper PerformanceEvaluation of a Novel 4D Trajectory Prediction Model for Civil Aircraft, Porretta proposed an aircraft performance model that comprehensively considers wind speed, the aircraft's lateral braking force and speed estimation. However, the model parameters required by these methods, such as weather forecasts, surface control intentions, flight plan information, etc., are difficult to obtain in actual applications. Therefore, in the absence of these model parameters, the target cannot be accurately modeled, which will lead to prediction The results are inaccurate.

In terms of parameter optimal estimation, the most classic algorithm is to estimate the target state based on Kalman filter to predict the trajectory of the aircraft. In her paper on tracking moving targets on airport surfaces based on the IMM algorithm, Gong Shuli proposed an algorithm based on interactive multi-model combined with unscented Kalman filtering. This algorithm models the movement process of the aircraft and performs trajectory prediction calculations. Tang Xinmin proposed a trajectory prediction algorithm based on hybrid system theory in his paper Conflict-free 4D trajectory prediction based on hybrid system theory. This algorithm constructs the parameter evolution of the aircraft dynamics model based on the movement characteristics of the aircraft in different flight segments. model, and built a state transition model to model the switching between different flight segments, and completed the aircraft trajectory prediction by adjusting the corresponding parameters. However, this type of algorithm has low execution efficiency, and cannot accurately model the target when the target's motion state is unknown, resulting in large errors in predicted track. In terms of machine learning models, Ma Yong proposed a precise track prediction method based on data mining in his paper Research on the Precision Prediction Method of Four-Dimensional Tracks Based on Data Mining. This method first clusters historical tracks and then calculates The dense trajectories of each cluster are combined with the hidden Markov model to achieve map matching of the air transportation network to complete accurate trajectory prediction. This type of algorithm is more accurate in forecasting than classic forecasting methods, but still suffers from the problem of shorter forecasting time.

In summary, existing target trajectory prediction technologies have the disadvantages of large errors, single prediction models, and short prediction duration, resulting in poor accuracy of predicted trajectory and affecting flight safety.

Contents of the invention

The purpose of the present invention is to propose a cyclic prediction method based on resampling to reduce the prediction error of maneuvering target tracks, increase the prediction time, and improve the accuracy of track prediction in view of the above-mentioned shortcomings of the existing technology.

The technical solution of the present invention is to preprocess the target historical track data; construct and train a neural network model; use a loop strategy to generate part of the historical track data, and use the trained network model parameters to calculate it; and calculate the calculation results Perform smoothing filtering to obtain the final predicted track.

According to the above ideas, the present invention is based on the track cycle prediction method under resampling, which is characterized by including the following:

(1) Filter and normalize the historical track data of maneuvering targets;

(2) Set the resampling period T' to 10 times the system sampling period T, sample the normalized track data into 10 resampled tracks as a network data set, and determine the input sample dimensions and labels sample dimensions;

(3) Construct a neural network consisting of four layers of bidirectional long short-term memory unit Bi-LSTM layer, Dropout layer, Dense layer, and activation layer;

(4) Set the maximum number of iterations to N and the batch size batch_size, send the network data set to the built network, and use the batch gradient descent method to iteratively train the parameters of the network. When the number of iterations reaches N, the training is completed network model;

(5) Predict the trajectory in the future:

(5a) Use a loop strategy to process the 10 resampled track data to generate 10 batches of data for predicting future tracks;

(5b) Calculate the 10 batches of data processed by the loop strategy sequentially by calling the trained neural network parameters to obtain 10 predicted tracks, and combine the 10 tracks in the order of the original time slots before resampling Predict the trajectory into one root;

(6) Use the smooth method to perform smoothing filtering on the predicted trajectory to obtain the final prediction result.

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

1) This invention uses the Bi-LSTM layer of the neural network to bidirectionally read the historical track of the target and learn its changing characteristics. When the target's motion state changes in the future, it can still predict the future track relatively accurately.

2) The present invention uses a resampling strategy to resample the historical track in different time slots, which can make the data set more balanced to reflect the change information of the historical track, reduce the prediction error of the neural network predicted track, and further improve Forecast accuracy.

3) The present invention uses a loop strategy to generate multiple batches of data for the resampled track data, and predicts the multiple batches of data in sequence. Compared with the traditional single prediction method, the present invention makes greater use of the data of the known track. Information can be used to obtain a predicted trajectory with a longer prediction time, provide a more accurate flight strategy for the target, and ensure the safety of the target flight.

Description of drawings

Figure 1 is a flow chart of the implementation of the present invention;

Figure 2 is a structural diagram of the neural network model used in the present invention.

Detailed ways

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

Referring to Figure 1, the present invention is based on the track cycle prediction method under resampling. The implementation steps are as follows:

Step 1: Generate track data set.

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

The target performs periodic motion with a cycle of 600s. The maneuvering state parameters within a single cycle are as shown in Table 1:

Table 1 Target maneuver information in some periods

Table 1 shows the target's maneuvering state changes in the first period (0 to 600s). In a single period, it lasts for 300s of constant speed movement, 150s of left turning movement, and 150s of right turning movement. The entire trajectory of the target has a total of five cycles, lasting 3000s. It is also assumed that the first 2400s of the target track are the known historical track, and the last 600s are the unknown future track.

The target is observed through the radar sensor observation platform, the observation sampling frequency is set to 1s, the radar delay is set to 1s, observation noise is added based on the radar angle and radius, and the noise size is set to [0.001°, 100m];

An interactive multi-model algorithm is used to fit the motion state of the above target, and the known track data of the first 2400s and the unknown track data of the last 600s are generated.

Step 2: Perform Kalman filtering on the known track information of the target 2400s before.

(2.1) Calculate the state prediction value of the target at the next moment k and error covariance P _k|k-1 :

Among them, k is the discrete time period, F is the target state transition matrix, T is the matrix transpose, is the state value at the current moment, P _k-1|k-1 is the error covariance at the current moment, Q is the process noise covariance matrix;

(2.2) Calculate track status update value and error covariance update value P _k|k :

P _k|k ＝[I-[P _k|k-1 H ^T (HP _k|k-1 H ^T +R) ^-1 ]H]P _k|k-1 ,

Among them, H is the observation matrix of the track, R is the measurement noise covariance matrix, Z _k is the radar measurement data at time k, and I is the unit matrix;

(2.3) Repeat (2.1) and (2.2) a total of 2400 times to obtain the filtered trajectory.

Step 3: Normalize the filtered track to obtain the normalized track data X _s :

Among them, _xi is the i-th time slot data of the filtered track.

Step 4: Sample the normalized track data X _s into 10 resampled tracks.

The specific implementation of this step is as follows:

Use the first resampled track to select the first time slot of the original track as the starting point, and perform resampling with a resampling interval T' of 10s;

Use the second resampled track to select the second time slot of the original track as the starting point, and perform resampling with a resampling interval T' of 10s;

By analogy, the 10th resampled track selects the 10th time slot of the original track as the starting point, and performs resampling with a resampling interval T' of 10s;

Each resampled track contains 300 sampling points, and the corresponding relationship between the time slots of the 10 resampled tracks and the time slots of the original track is obtained, as shown in Table 2:

Table 2 Correspondence table between resampled tracks and original track time slots

Step 5: Determine the network input sample dimensions and label sample dimensions.

The specific implementation of this step is as follows:

To determine the dimension of the network input sample: first determine the input sample sliding window length W as 120 and the number of features as 1; then perform sliding segmentation on the resampled 10 track data in sequence according to W, and splice the segmentation results. Obtain the network input sample with a sample batch size of 610 and a dimension of 610*120*1;

For the determination of the label sample dimension: Assume that the sample batch and number of features of the label sample are the same as the input sample, and the time steps of each batch of samples are 60, then a label sample with a dimension of 610*60*1 can be obtained;

The resampling track time slot corresponding to the input sample and label sample is as shown in Table 3:

Table 3 Correspondence table between input samples/label samples and track time slots

Step 6: Build the neural network model.

Referring to Figure 2, the neural network model constructed in this step consists of four layers: Bi-LSTM layer, Dropout layer, Dense layer, and activation layer from top to bottom. The functions and parameters of each layer are as follows:

The Bi-LSTM layer is used to extract the changing characteristics of the historical track data set, and its hidden node number units is 200;

The Dropout layer is used to prevent overfitting of the network during the training process, and its dropout rate dropout_ratio is 0.2;

The Dense layer is used to fit the label sample Y_train during network training, and its hidden node number units is 60;

The activation layer is used to enhance the adaptability of the network model to nonlinear data, and its activation function is a linear activation function.

Step 7, train the neural network model.

Existing methods for training neural networks include batch gradient descent, stochastic gradient descent, and mini gradient descent. This step uses but is not limited to the batch gradient descent method, which is implemented as follows:

(7.1) Set the data batch size batch_size to 64, divide the track data set into multiple small batches of data according to the data batch size, and send these small batches of data to the neural network in turn for single training;

(7.2) Set the network optimization algorithm to the adaptive moment estimation algorithm Adam, and optimize the network parameters by calculating and correcting the first-order moment and second-order moment of each round of training gradient;

(7.3) Set the maximum number of iterations of the network to 100, repeat (7.1) and (7.2) to reach the maximum number of iterations, and obtain the trained network model.

Step 8: Generate 10 batches of data for predicting future trajectories.

The specific implementation of this step is as follows:

(8.1) Calculate the prediction batch N of 10 batches of data according to the duration T _pre that needs to be predicted, the resampling period T' and the system sampling period T:

N=T _pre /T/T',

Among them, in this example, T _pre is 600s, T is 1s, T' is 10s, and the predicted batch N is 60 according to the formula;

(8.2) Use a loop strategy to select 60 batches of test data for 10 resampled tracks to generate 10 batches of data:

For the first batch of test data of the first batch of data, select the track data of the 62nd to 181st time slots of the first resampled track;

For the test data of the second batch of the first batch of data, select the track data of the 63rd to 182nd time slots of the first resampled track;

By analogy, for the test data of the 60th batch of the first batch of data, the track data of the 121st to 240th time slots of the first resampled track are selected;

For the first batch of test data of the second batch of data, select the track data of the 62nd to 181st time slots of the second resampled track;

For the test data of the second batch of data, select the track data of the 63rd to 182nd time slots of the second resampled track;

By analogy, for the test data of the 60th batch of the 10th batch, the track data of the 121st to 240th time slots of the 10th resampled track are selected.

The corresponding relationship between 60 batches of test data and the 10 resampled track time slots is shown in Table 4:

Table 4. Track time slot correspondence table after network input resampling

Step 9: Calculate the 10 batches of data processed by the loop strategy sequentially by calling the trained neural network parameters to obtain 10 predicted tracks.

(9.1) Calculate the 60 batches of test data of the first batch of data obtained in (8.2) in sequence:

Call the trained network parameters to calculate the first batch of test data of the first batch of data, and obtain the trajectory of time slots 182 to 241;

Call the trained network parameters to calculate the second batch of test data of the first batch of data, and obtain the trajectory of time slots 183 to 242;

By analogy, the trained network parameters are called to calculate the 60th batch of test data of the first batch of data, and the trajectory of time slots 241 to 300 is obtained.

The corresponding relationship between the 60 batches of prediction results obtained above and the resampled track time slots is shown in Table 5:

Table 5 Correspondence table of 60 batches of prediction results and track time slots after resampling

(9.2) Select the last time slot point of 60 batches of prediction results in sequence and splice it into a predicted track containing 60 time slot points, that is, select the last time slot point 241 of the prediction result in the first prediction batch, and then select the last time slot point 241 of the prediction result in the first prediction batch. 2 Select the last time slot point 242 of the prediction result in the 2 prediction batch, and so on, select the last time slot point 300 of the prediction result in the 60th prediction batch, and splice these data in sequence to form 60 time slot points The track data is the predicted track of the first batch of data.

(9.3) Use the methods (9.1) and (9.2) to predict the 2nd to 9th batches of data to obtain the predicted tracks of the 2nd to 9th batches of data,

(9.4) Combine the 10 predicted tracks obtained from (9.1)-(9.3) into one predicted track in the order of the original time slots before resampling.

Step 10: Use the smooth method to perform smoothing filtering on the predicted trajectory to obtain the final prediction result.

The formula of smoothing filter is as follows:

yy(n)=(y(1)+y(2)+y(3)+...+y(n))/n,

Among them, y(n) represents the value of the nth element before smoothing, yy(n) represents the value of the nth element after smoothing, and the smoothed predicted trajectory is the final prediction result.

The effect of the present invention can be further illustrated through the following simulation experiments.

1. Simulation conditions:

The neural network model uses the Keras framework of Python 3.6 as the simulation platform;

The track generation part and Kalman filtering part use MATLAB 2018a as the simulation platform;

2. Simulation content:

Under the above conditions, the present invention and the existing two trajectory prediction algorithms are used to predict the target's trajectory for the next 600s for a long time, and the prediction results are obtained, as shown in Table 6:

Table 6 Comparison of results of different trajectory prediction algorithms

It can be seen from Table 6 that compared with the existing algorithm, the track cycle prediction method based on resampling proposed by the present invention has lower error when predicting the track, the prediction time is longer, and can provide more accurate targets for the target. Accurate flight strategy to ensure target flight safety.

The above description is only a specific example of the present invention and does not constitute any limitation on the present invention. Obviously, for professionals in the field, after understanding the content and principles of the present invention, it is possible to make decisions without departing from the principles and structural ideas of the present invention. Under the circumstances, various modifications and changes in form and details are made, but these modifications and changes based on the idea of the present invention are still within the scope of the claims of the present invention.

Claims (10)

1. A track cycle prediction method based on resampling, which is characterized by the following: (1) Filter and normalize the historical track data of maneuvering targets; (2) Set the resampling period T' to 10 times the system sampling period T, sample the normalized track data into 10 resampled tracks as a network data set, and determine the input sample dimensions and labels sample dimensions; (3) Construct a neural network consisting of four layers of bidirectional long short-term memory unit Bi-LSTM layer, Dropout layer, Dense layer, and activation layer; (4) Set the maximum number of iterations to N and the batch size batch_size, send the network data set to the built network, and use the batch gradient descent method to iteratively train the parameters of the network. When the number of iterations reaches N, the training is completed network model; (5) Predict the trajectory in the future: (5a) Use a loop strategy to process the 10 resampled track data to generate 10 batches of data for predicting future tracks; (5b) Calculate the 10 batches of data processed by the loop strategy sequentially by calling the trained neural network parameters to obtain 10 predicted tracks, and combine them into one according to the time slot sequence of the original tracks before resampling. root predicted trajectory; (6) Use the smooth method to perform smoothing filtering on the predicted trajectory to obtain the final prediction result. 2. The method according to claim 1, characterized in that filtering the track data in (1) uses a Kalman filter algorithm to perform state prediction and state update on the historical track data to eliminate system variables. To measure noise, the implementation is as follows: (1a) Calculate the state prediction value of the target at the next time k and error covariance P _k|k-1 : Among them, k is the discrete time period, F is the target state transition matrix, and Q is the system process noise; (1b) Calculate track status update value and error covariance update value P _k|k : Pk _|k ＝ _{[ IGkHk} ]Pk _|k-1 Among them, H is the observation matrix of the track, R _k is the noise covariance matrix, T is the matrix transpose, and Z is the radar measurement data; (1c) Perform iterative calculations on (1a) and (1b), the number of iterations is the number of time slot points of the historical track, and obtain the filtered target track. 3. The method according to claim 1, characterized in that, in (1), the track data is normalized, and the formula is as follows: Among them, _xi is the i-th time slot data of the filtered track, and X _scaled is the normalized data. 4. The method according to claim 1, characterized in that, in (2), the normalized track data is sampled into 10 resampled tracks, which is implemented as follows: The first resampled track selects the first time slot of the original track as the starting point, and performs resampling with a sampling interval of T'; The second resampled track selects the second time slot of the original track as the starting point and performs resampling with a sampling interval of T'; By analogy, the 10th resampled track selects the 10th time slot of the original track as the starting point, and performs resampling with a sampling interval of T', so that the original track becomes 10 resampled tracks. 5. The method according to claim 1, characterized in that (2) determines the network input sample dimension and the label sample dimension, and is implemented as follows: To determine the dimension of the network input sample, first determine the input sample sliding window length W as 120 and the number of features as 1, and then perform sliding segmentation on the resampled 10 track data in sequence according to W, and splice the segmentation results. Obtain the network input sample with a sample batch size of 610 and a dimension of 610*120*1; For the determination of the dimension of the label sample, assuming that the sample batch and number of features of the label sample are the same as the input sample, and assuming the number of time steps of each batch of samples is 60, a label sample with a dimension of 610*60*1 can be obtained. 6. The method according to claim 1, characterized in that the neural network constructed in (3) has the following functions and parameters: The Bi-LSTM layer is used to extract the changing characteristics of the historical track data set, and its hidden node number units is 200; The Dropout layer is used to prevent overfitting of the network during the training process, and its dropout rate dropout_ratio is 0.2; The Dense layer is used to fit the label sample Y_train during network training, and its hidden node number units is 60; The activation layer is used to enhance the adaptability of the network model to nonlinear data, and its activation function is a linear activation function. 7. The method according to claim 1, characterized in that, in (4), the mini-batch gradient descent method is used to iteratively train the parameters of the neural network, which is implemented as follows: (4a) Set the data batch size batch_size to 64, divide the track data set into multiple small batches of data according to the data batch size, and send these small batches of data to the neural network in turn for single training; (4b) Set the network optimization algorithm to the adaptive moment estimation algorithm Adam, and optimize the network parameters by calculating and correcting the first-order moment and second-order moment of each round of training gradient; (4c) Set the maximum number of iterations of the network to 100, repeat (4a) and (4b) a total of 100 times, and obtain the trained network model. 8. The method according to claim 1, characterized in that (5a) generates 10 batches of data for predicting future trajectories, which is implemented as follows: (5a1) Calculate the prediction batch N for each batch of 10 batches of data according to the duration T _pre that needs to be predicted, the resampling period T' and the system sampling period T: N=T _pre /T/T'; (5a2) Use a cyclic strategy to select N batches of network inputs for 10 resampled tracks to generate 10 batches of data: For the first batch of test data of the first batch of data, select the track data of the 62nd to 181st time slots of the first resampled track; For the test data of the second batch of the first batch of data, select the track data of the 63rd to 182nd time slots of the first resampled track; By analogy, the Nth batch of test data of the first batch of data selects the track data of the 121st to 240th time slots of the first resampled track; For the first batch of test data of the second batch of data, select the track data of the 62nd to 181st time slots of the second resampled track; For the test data of the second batch of data, select the track data of the 63rd to 182nd time slots of the second resampled track; By analogy, the test data of the Nth batch of the 10th batch are selected from the track data of the 121st to 240th time slots of the 10th resampled track. 9. The method according to claim 1, characterized in that in (5b), the 10 batches of data are calculated sequentially by calling the trained neural network parameters, and the implementation is as follows: (5b1) Call the trained network parameters to calculate the N batches of test data of the first batch of data obtained in (5a) in sequence, and obtain N batches of prediction results; (5b2) Select the last time slot point of N batches of prediction results in sequence and splice them into a predicted track containing N time slot points, which is the predicted track of the first batch of data; (5b3) Use the methods (5b1) and (5b2) to predict the 2nd to 9th batches of data to obtain the predicted tracks of the 2nd to 9th batches of data, and put the 10 predicted tracks in the order of the original time slots before resampling. Combined into a predicted trajectory. 10. The method according to claim 1, characterized in that in (6), the smooth method is used to perform smoothing filtering on the predicted trajectory, and the formula is as follows: yy(n)＝(y(1)+y(2)+y(3)+...+y(n))/n Among them, y(n) represents the value of the nth element before smoothing, yy(n) represents the value of the nth element after smoothing, and the smoothed predicted trajectory is the final prediction result.