CN114048889B - Aircraft trajectory prediction method based on long-term and short-term memory network - Google Patents

Abstract

The invention relates to the fields of air combat environment, data processing, deep learning and the like, and provides a method for realizing aircraft trajectory prediction by using a long-short term memory network (LSTM) under the uncertain perception condition. Therefore, the technical scheme adopted by the invention is that a method for predicting the aircraft track based on a long-term and short-term memory network utilizes Kalman filtering to eliminate the noise interference carried by the sensor characteristic vector; for the directly acquired state parameters, data preprocessing is carried out, including down-sampling, invalid value elimination and missing value complementation, in addition, in order to improve the calculation stability, the data is subjected to normalization processing, and the value range of the input data is included in a [0,1] interval; and constructing a locus prediction model based on the LSTM, defining input and output of the network, and carrying out supervision training on the network. The method is mainly applied to occasions for predicting the flight path of the unmanned aerial vehicle.

Description

Aircraft trajectory prediction method based on long-term and short-term memory network

Technical Field

The invention relates to the fields of air combat environment, data processing, deep learning and the like, and solves the problem of predicting flight trajectories of aircrafts under uncertain perception conditions. In particular to a method for predicting the trajectory of an aircraft based on a long-term and short-term memory network.

Background

Under the current international environment, the air force is the embodiment of the whole fighting force of a country. Under the actual air battle environment, a driver needs to master the flight state of an enemy plane in real time according to the real-time data information of the airborne sensor. If the future state information of the enemy plane, including the position and the possible flight track which may appear in the future, can be predicted in advance according to the existing sensor state parameter information, the method is beneficial to the party to carry out the fighting strategies such as interception, attack, escape and the like in advance, and the party's winning rate is improved. Therefore, the position of the enemy aircraft at the next moment can be predicted dynamically and accurately through the known information, and the method has important strategic significance. However, the conventional trajectory prediction method has the problems of serious model simplification and few consideration factors, and is difficult to process the mutually coupled state information, so that a more accurate prediction result is difficult to give. As an emerging prediction method, the neural network has strong fitting capability to describe complex nonlinear relations compared with the traditional prediction method.

The track prediction means that a prediction model is built by using the existing historical track data information to obtain the position point of the future time. At present, three methods are mainly used for constructing a track prediction model:

one is based on association rules ^[1] The trajectory prediction of (1). And constructing an association rule by mining frequently-occurring items, and predicting the track by using a sequence matching method. The track prediction by using the association rule is mainly divided into two subtasks: the first part is the mining of a frequent item set, and position points with high occurrence frequency are found out from the existing historical data and are called as frequent items, and the formed set is called as the frequent item set; and the second part is to generate an association rule, calculate the probability that one position appears and the other position also appears in the frequent item set in the first part, wherein the probability is called confidence, and when the current position is input into the rule base, the position with the highest confidence is output as a track prediction result. In the association rule algorithm, Apriori is typical ^[2] The algorithm can rapidly and accurately dig out the association rule, but the database needs to be repeatedly scanned, a large number of useless feature sets are generated, and the calculation complexity is high. The Prefix Span algorithm calculates frequent items and association rules according to the sequence of the track, and uses the Prefix projection technology to find the subsequent position of a certain position according to the time sequence to form a frequent item set, so that the frequent item set has certain continuity. Document [3]Combining the FP-Tree algorithm and the Prefix Span algorithm, excavating frequent tracks, matching the existing tracks with a motion rule base, and establishing a probability model of the position of an object. The research only considers historical track information, and although the mined frequent items have certain continuity, certain positions can be skipped in the middle, so that the accuracy of the prediction result is not high.

The other is Markov model based trajectory prediction. Calculating the probability from a certain position point to other position points by constructing a state transition matrix, and inputting the current position into the constructed positionAnd in the matrix, determining the next position according to the maximum probability so as to obtain a prediction result. The first order Markov model has only one position transition probability matrix, and Yang J is used for obtaining more comprehensive information ^[4] A high-order Markov model is designed, the position of the next moment is predicted by using the state information of n positions, and the prediction accuracy is improved. Qiao S J ^[5] The hidden Markov-based track prediction algorithm HMTP improves the problem that the speed change of an object in motion is fast and difficult to predict, and can predict the continuous track of the object. The research is based on the assumption that the position information at the current moment is only related to the previous moment, the obtained local optimal solution is obtained, and the high-order Markov calculation has high complexity and is not suitable for the requirement of real-time prediction.

And thirdly, predicting the track based on the neural network. The neural network is generally composed of an input layer, a hidden layer and an output layer, the output of the forward propagation calculation network and the error between the backward propagation calculation output and a true value are output, so that the network weight is updated, and a prediction model with high fitting degree with the true model can be obtained through multiple parameter updates. Payeur ^[6] The other people propose that the motion track of the robot is predicted by utilizing an Artificial Neural Network (ANN), six nearest measurement quantities of the robot coordinate are used as the input of the network, and the future track position is predicted through a plurality of neurons; huang ^[7] The model is provided with a Deep Belief layer at the bottom for unsupervised training and a multi-task learning layer at the top for supervised learning; billow ^[8] The method comprises the steps of carrying out trajectory prediction on a fighter by utilizing an Elman neural network, selecting position point values in three directions as position vectors, sampling the flight trajectory of the fighter every 0.25 second, predicting the next position point by using the first five position points, constructing the Elman neural network, and training to obtain a prediction result; thrice of the prince thrifty ^[9] On the basis of the Elman neural network, the particle swarm algorithm and the gradient algorithm are utilized to optimize the network weight, the landing track of the unmanned aerial vehicle is well predicted, and the prediction speed and the prediction accuracy of the model are improvedDegree; money with Kui, etc ^[10] When the flight track of the airplane is predicted, firstly, a target track group is subjected to cluster analysis by using a self-adaptive K-means algorithm, a specific activity area of a certain target is found out, then, a BP neural network is used for training the track group, a prediction model is established, and the prediction of the track is completed; yangyongnong (Yangyongnong) ^[11] When the trajectory of the unmanned aerial vehicle is predicted, a Bi-LSTM-based prediction model is provided, parameters such as position, attitude, distance and the like are used as input of the model, the position at the next moment is used as output, the prediction of the flight trajectory of the unmanned aerial vehicle is completed by a training network, and experiments show that the trajectory prediction result of the method is more accurate than that of the trajectory prediction based on the Elman neural network.

The trajectory prediction of the aircraft is essentially the prediction of time series data, the flight trajectory has high uncertainty due to the complex and changeable environment, the traditional prediction method uses small data amount, the mutual influence of the position and the attitude of the aircraft is not considered, the flight rule of the aircraft influenced by multiple factors is difficult to accurately learn, and therefore a more intelligent prediction model is needed. The idea of reference [11] adopts a datamation method to convert the trajectory prediction problem into the prediction problem of time series data, and starts from flight data only, does not consider the physical model of the aircraft, selects a deep learning prediction model based on a long-short term memory network (LSTM), and simultaneously, when selecting the network input characteristic quantity, not only considers the state parameters of the opposite side obtained by the sensor, but also adds own state parameters, so that the flight trajectory of the aircraft influenced by multiple factors is fitted to the maximum degree by the prediction model.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention aims to provide a method for realizing aircraft trajectory prediction by using a long-short term memory network (LSTM) under the uncertain perception condition. Therefore, the technical scheme adopted by the invention is that a method for predicting the aircraft track based on the long-term and short-term memory network is used for eliminating noise interference carried by the sensor characteristic vector by using Kalman filtering; for the directly acquired state parameters, data preprocessing is carried out, including down-sampling, invalid value elimination and missing value complementation, in addition, in order to improve the calculation stability, the data is subjected to normalization processing, and the value range of the input data is included in a [0,1] interval; and constructing a locus prediction model based on the LSTM, defining input and output of the network, and carrying out supervision training on the network.

The method comprises the following specific steps:

the method comprises the steps that firstly, original state information is obtained from an on-board sensor, twenty-three-dimensional characteristic quantities are selected as the input of a prediction model;

the second step is to preprocess the data: firstly, the original sensor parameters contain certain noise interference, and a Kalman filtering method is selected to eliminate the noise interference; secondly, in order to improve the data utilization rate and reduce the calculation cost, the data processing method comprises the following steps of 1: 5, performing down-sampling, and performing normalization processing on the data; finally, rejecting the invalid value, complementing the missing value and finishing the data preprocessing work; then, input and tag value selection are carried out: the segmentation of the data set is completed by adopting a sliding window method, the input of the network is obtained by iterative selection of original data, from 0 to the first time _ step, the time _ step selected in the training is 20, namely, from 0 to 19 is the first input, from 1 to 20 is the second input, and the like; the dimensionality of input data is (N,20,23), wherein N depends on the size of a sample set, 20 refers to a time _ step value of an LSTM, 23 is the dimensionality of a feature vector, label data are needed by a training network, and the purpose of the model is to predict the track of the airplane in one step in the future, so that the longitude and latitude heights of the airplane carrier are selected to be labels from 1 to 20 for the first input from 0 to 19, the longitude and latitude heights of the airplane carrier are selected to be labels from 2 to 21 for the second input from 1 to 20, and the like;

and thirdly, constructing a track prediction model based on the LSTM, wherein the constructed model comprises two LSTM layers, a Dropout layer is added behind each LSTM layer, the phenomenon of overfitting is avoided, and the dimension is converted into the required output dimension after the output characteristic quantity passes through a full connection layer.

The detailed steps are as follows:

firstly, acquiring required characteristic data from a hostile and antagonistic simulation platform of an enemy aircraft, and summarizing the aircraft track obtained by considering the influence of various factors such as the aircraft track, battlefield environment and enemy motivation to have the three characteristics of continuity, time sequence and interactivity:

continuity means that the unmanned aerial vehicle track is continuously changed, but not discontinuous;

the time sequence is that the track data is time-nature, and the position of the next moment is related to the position of the previous moment, so the track data is a time sequence data in nature;

interactivity means a complex process of dynamic change among multiple airplanes in an actual environment, the maneuver of one airplane can influence the maneuver of another airplane, and the position of one airplane can also influence the position of another airplane;

therefore, when the trajectory prediction is performed, in addition to the position, the posture and the speed of the enemy plane, the 23-dimensional feature vectors of the two planes are considered due to the interactivity:

longitude, latitude, altitude, pitch angle, roll angle, course angle, azimuth angle speed, pitch angle speed, north direction speed, sky direction speed, east direction speed, north direction acceleration, sky direction acceleration, east direction acceleration, horizontal entering angle, relative distance change rate, relative altitude, radar state identification result, radial speed, true course and ground speed;

secondly, after the feature vector is obtained, since the raw data obtained from the sensor cannot be directly used as network input, it needs to be preprocessed, according to 1: 5, carrying out data downsampling on data at sampling intervals, removing invalid values in the data, and complementing missing values by using a mean filling method, in addition, because the numerical values of part of used parameters are large, in order to improve the calculation stability, carrying out normalization processing on the data, and enabling the value range of input data to be included in a [0,1] interval, wherein the normalization formula is as follows:

wherein X is the actual value of a certain characteristic quantity, and X is _max ,X _min Respectively the maximum value and the minimum value of X in all data, and Y is the result after normalization;

and (3) performing inverse normalization on the prediction result by using the trained prediction model so as to perform error comparison analysis with the actual value, wherein the inverse normalization formula is as follows:

X＝(X _max -X _min )Y+X _min (2)

after the data processing is completed, a normalized sample data set available for the neural network is obtained, at this time, input and output of the LSTM network need to be constructed according to actual requirements, when label data are selected, longitude, latitude and height of the next moment are taken as feature labels of the current moment, segmentation of the data set is completed by adopting a sliding window method, input of the network is obtained by iterative selection of original data, the time _ step is from 0 to the first time _ step, the time _ step selected by the training is 20, namely from 0 to 19 is the first input, from 1 to 20 is the second input, and the like; the dimensionality of input data is (N,20,23), wherein N depends on the size of a sample set, 20 refers to a time _ step value of an LSTM, 23 refers to the dimensionality of a feature vector, the label data is needed by a training network, and the object of the model is to predict the future trajectory of the airplane one step, so that the latitude and longitude heights of an airplane carrier are selected as labels at the input of the first 0-19, the longitude and latitude heights of the airplane carrier are selected as labels at the input of the second 1-20, and the like, so that the dimensionality of output data is (N,20,3), wherein N depends on the size of the sample set, 20 refers to the time _ step value of the LSTM, and 3 corresponds to the output dimensionality, namely the longitude, the latitude and the height of the airplane carrier;

and finally, constructing an LSTM network, using the LSTM as a main body part of the model, learning the relation between the input track related characteristic quantity and a label of the next time position by constructing a multi-layer network structure, predicting the position which is possibly reached in the future, wherein the prediction model mainly comprises an LSTM layer, a Dropout layer and a full connection layer, and selecting two layers of LSTMs to build the model: the LSTM network of the first layer takes twenty-three-dimensional characteristic quantity as input, and the second layer takes the output of the first layer

As input, second layer output

After passing through a full connection layer, obtaining a final network output Y, adding a Dropout layer in each LSTM network, controlling the node weight of a hidden layer by using the Dropout layer, avoiding certain track characteristics from being effective only under fixed combination, consciously enabling the network to learn the universality of tracks, and converting the dimensionality into the required output latitude after passing through the full connection layer on the characteristic quantity output by the LSTM;

after the network structure and input and output are determined, training the network by using a training sample set, and selecting a mean square error function shown as a formula (3) as a loss function:

wherein N represents the number of batch samples in one training process, Y _pred Predicted values representing neural network outputs, Y _i Representing the corresponding real value, the neural network training is the process of updating the weight, the optimization goal of the network is to make E approach to 0, when the network training is carried out, an adaptive moment estimation method (Adam) is selected, and the adaptive moment estimation method comprises the following steps:

firstly, considering that in the traditional back propagation algorithm, the updating direction of the weight only depends on the gradient obtained by the current sample, so the concept of momentum (moment) in physics is used for reference, namely, the direction updated before is kept to a certain extent when the weight is updated, and the final updating direction is obtained by adding the gradient of the current sample, namely, the final updating direction is obtained

Where s is called momentum, and is also an estimate of the first moment of the gradient, β ₁ Referred to as the first order momentum decay coefficient;

where v is called the velocity quantity, and is also an estimate of the second moment of the gradient, β ₂ Called second order momentum decay coefficient, typically taken as beta ₂ ＝0.999；

Further, to realize adaptive adjustment of the learning rate, i.e. to use a larger epsilon update for the weight with a slower update and a smaller epsilon update for the weight with a faster update, the learning rate epsilon is adjusted:

δ is a constant to prevent the denominator from being zero, and δ is generally taken to be 10 ^-8 ；

On the basis of the above, the first and second moment estimates of the gradient are corrected unbiased:

and finally, obtaining a weight updating formula based on the Adam algorithm:

according to the steps, the training of the LSTM network is completed, the trained model is stored locally, the sample mean value and variance parameters are stored in a txt file format, and the local files are directly called when a new sample to be tested is tested;

when model training and testing are performed, the values of network structure parameters are shown in the following table:

TABLE 1 LSTM network parameters

Inputting new data to be tested for the trained prediction model, and verifying the model prediction accuracy;

when a sample data set is selected, in a simulation countermeasure environment, each time a simulation countermeasure is completed, a group of experimental data, namely a track curve, is obtained, the condition that characteristic points of a single curve are too few and are not beneficial to learning of a neural network is considered, therefore, multiple experiments are carried out, tracks are spliced, time sequence data containing 10600 points are finally obtained, each point contains twenty-three-dimensional characteristic quantity, after dimensionality reduction, a sample set with dimensionality (21200,23) is obtained, a training set and a testing set are divided according to the proportion of 0.75:0.25, and the sample dimensionality of the testing set is (5300, 23);

in order to prevent the model from being over-fitted within the set epoch range, in addition to adding a Dropout layer when designing the model, sample data is processed, for training data, a part of the training data is selected as verification data (valid data), the proportion of the verification set data to the training set data is defined as valid _ data _ rate being 0.15, a endurance value probability being 5 is set, if the iteration number of the verification set exceeds 5 times and the training precision and loss are not improved, the training is stopped in advance, which is called as Early training Stopping (Early Stopping), and by adding the verification set, the over-fitting condition can be avoided, and the model selection is completed.

The invention has the characteristics and beneficial effects that:

when the aircraft trajectory is predicted, due to the complexity of the environment of the aircraft, the high maneuverability of the aircraft and the interactivity among the aircraft, the traditional prediction method is difficult to obtain an accurate prediction result. The invention adopts the deep learning prediction model based on the LSTM network, can process a large amount of aircraft data information, fits the motion model of the aircraft from the aircraft data information, and further provides a high-precision prediction result. Meanwhile, the model is purely from the data perspective, the influence of the aircraft physical model is omitted, and the complexity of model construction is simplified to a certain extent. Twenty-three-dimensional characteristic quantities related to track prediction are selected from the state parameters of the three-hundred multi-dimensional sensor, and filtering, down-sampling, missing value complementing, abnormal value removing and normalization are realized through a data preprocessing module; selecting input quantity and label value by using a sliding window method; after the model is built, the network is supervised and learned by using the input quantity and the label data, network parameters are optimized, and new data to be tested are input to the trained network, so that a prediction result of a track at the next moment can be obtained.

The invention mainly has the following characteristics and advantages:

(1) according to the twenty-three-dimensional characteristic quantities of the two aircrafts, the influence factors of the aircraft track of the aircraft can be described to the maximum extent, and a prediction model is fitted.

(2) And dividing the input quantity and the label value, taking the longitude, the latitude and the altitude of the next moment as the label value of the current moment, and directly obtaining a prediction result when new data is input.

(3) A prediction model formed by multiple layers of LSTMs can fit complex nonlinear tracks, and meanwhile, the phenomenon of overfitting is avoided.

Description of the drawings:

FIG. 1 is a flow chart of trajectory prediction.

FIG. 2 is a schematic diagram of input quantity and tag value values.

FIG. 3 is a diagram of a network architecture of a trajectory prediction model.

Figure 4 training loss curves.

FIG. 5 aircraft trajectory prediction results.

Detailed Description

In view of the foregoing background and problems, the present invention is directed to a method for implementing aircraft trajectory prediction using Long Short Term Memory (LSTM) under uncertain sensing conditions. Eliminating noise interference carried by the sensor characteristic vector by using Kalman filtering; for the directly acquired state parameters, data preprocessing is carried out, including down-sampling, invalid value elimination and missing value complementation, in addition, in order to improve the calculation stability, the data is subjected to normalization processing, and the value range of the input data is included in a [0,1] interval; and constructing a track prediction model based on the LSTM, defining input and output of the network, and performing supervision training on the network.

The invention has the following functions and characteristics:

(1) according to the characteristic that the track of the aircraft has interactivity in the actual air combat environment, the state parameter information of the twenty-three-dimensional double-side aircraft is selected as the input vector of the network, and the longitude vector, the latitude vector and the altitude vector are selected as the output of the network.

(2) The invention provides a network input and label determination method, which utilizes a sliding window method to complete the segmentation of a data set by taking the longitude, the latitude and the height of the next moment as the characteristic label of the current moment.

(3) The invention provides an aircraft trajectory prediction model based on an LSTM network, which comprises an LSTM layer, a Dropout layer and a full connection layer. A Dropout layer is added in the LSTM network of each layer, and the Dropout layer is used for controlling the node weight of the hidden layer, so that certain track characteristics are prevented from being effective only under a fixed combination, and the network is consciously led to learn the common commonality of the tracks.

The technical scheme of the invention is as follows:

the main purpose of the invention is to realize one-step prediction of the aircraft track, and the general flow is shown in figure 1.

The first step is to obtain original state information from onboard sensors, wherein the number of the sensors available for the aircraft is usually more than three hundred, and the sensors are selected according to needs in actual use. In this invention, twenty-three-dimensional feature quantities were selected in total as input to the prediction model, as shown in table 1.

The second step is to preprocess the data. Firstly, the original sensor parameters contain certain noise interference, and a Kalman filtering method is selected to eliminate the noise interference; secondly, in order to improve the data utilization rate and reduce the calculation cost, the data processing method comprises the following steps of 1: 5, performing down-sampling, and performing normalization processing on the data; and finally, removing the invalid value, complementing the missing value and finishing the data preprocessing work. Then, the input and tag values are selected according to the method shown in fig. 2.

And the third step is to construct a locus prediction model based on the LSTM, wherein the constructed model comprises two LSTM layers as shown in figure 3, and a Dropout layer is added behind each LSTM layer, so that the overfitting phenomenon is avoided. And after the output characteristic quantity passes through a full connection layer, converting the dimension into the required output dimension.

The invention will be further explained with reference to the drawings.

The aircraft track can be regarded as a series of time series data, so that a data idea can be adopted to convert the track prediction problem into a time series data prediction problem, and a proper input sample is selected by utilizing the strong data fitting capacity of the LSTM network to obtain a one-step prediction result of the aircraft track.

When predicting the aircraft trajectory, the process is performed according to the flow shown in fig. 1. The method mainly comprises the following four steps: (1) extracting effective data in the countermeasure simulation, and establishing a training sample data set; (2) pretreating a sample; (3) constructing network input and output, establishing a neural network, performing supervised training on the network by using the input and output, adjusting network structure parameters according to a training result, inputting a new sample to be tested by using the trained network, obtaining a prediction result, and comparing the prediction result with a true value.

First, the required feature data is obtained from the enemy-me aircraft countermeasure simulation platform. Considering that the track of the aircraft is influenced by various factors such as self, battlefield environment, enemy motivation and the like, the track of the aircraft is induced to have continuity, chronology and interactivity according to expert experience ^[12] Three characteristics.

Continuity means that the unmanned aerial vehicle trajectory is continuously changing, not discontinuous.

Chronology means that the track data is time-nature, and the position of the next time is related to the position of the previous time, so the track data is essentially time-series data.

Interactivity means a complex process of dynamic change among multiple airplanes in an actual environment, the maneuver of one airplane can affect the maneuver of another airplane, and the position of one airplane can also affect the position of another airplane.

Therefore, in addition to the position, attitude and speed of the enemy plane, twenty-three-dimensional characteristics including the relative distance, relative height and horizontal entrance angle of the two planes are considered for the interactivity in the track prediction, as shown in table 2.

TABLE 2 neural network input feature quantities

Secondly, after a total of twenty-three-dimensional feature vectors are obtained, since raw data acquired from a sensor cannot be directly used as network input, it needs to be preprocessed. According to the following steps of 1: 5, carrying out data downsampling on data at sampling intervals, removing invalid values in the data, and complementing missing values by using a mean filling method, in addition, because the numerical values of part of used parameters are large, in order to improve the calculation stability, carrying out normalization processing on the data, and enabling the value range of input data to be included in a [0,1] interval, wherein the normalization formula is as follows:

wherein X is the actual value of a certain characteristic quantity, and X is _max ,X _min The maximum and minimum values of X in all data, respectively, and Y is the normalized result.

And (3) performing inverse normalization on the prediction result by using the trained prediction model so as to perform error comparison analysis with an actual value, wherein the inverse normalization formula is as follows:

X＝(X _max -X _min )Y+X _min (2)

after the data processing is completed, a normalized sample data set available for the neural network is obtained, and at this time, the input and output of the LSTM network need to be constructed according to actual requirements. The ultimate goal here is to predict the longitude, latitude, and altitude of the aircraft at the next time from the twenty-three-dimensional feature quantities at the current time, and therefore, when the tag data is selected, the longitude, latitude, and altitude at the next time are taken as the feature tags at the current time. The segmentation of the data set is done using a sliding window method, as shown in fig. 2. The meaning of it is: the input of the network is obtained by iterative selection of the original data, starting from 0 to the first time _ step, since the time _ step selected for this training is 20, i.e. from 0 to 19 is the first input, 1 to 20 is the second input, and so on. The dimension of the input data is therefore (N,20,23), where N depends on the size of the sample set, 20 refers to the time step value of LSTM, and 23 is the dimension of the feature vector. The training network needs to use label data, and because the model aims to predict the future track of the airplane, 1-20 longitude and latitude heights of the airplane are selected as labels for the first input from 0 to 19, 2-21 longitude and latitude heights of the airplane are selected as labels for the second input from 1 to 20, and the like. The dimensionality of the output data is thus (N,20,3), where N depends on the size of the sample set, 20 refers to the time step value of LSTM, and 3 corresponds to the dimensionality of the output, i.e., the carrier longitude, latitude, height.

Finally, the LSTM network is constructed. The LSTM is used as a main part of the model, and the relation between the input track-related characteristic quantity and the label of the position at the next moment is learned by constructing a multi-layer network structure, so that the position which is possibly reached in the future is predicted. The prediction model mainly comprises an LSTM layer, a Dropout layer and a full connection layer. Two layers of LSTM were selected for model building herein, as shown in fig. 3: the LSTM network of the first layer takes twenty-three-dimensional characteristic quantity as input, and the second layer takes the output of the first layer

As input, second layer output

After a full connection layer, the final network output Y is obtained. Since the number of layers and neurons in the network structure is large, to prevent the "overfitting" phenomenon, Hinton proposed the use of Dropout method, which is trainedIn the process, some neurons for feature learning are randomly omitted, and overfitting caused by complex mutual adaptation generated on a training set due to the fact that only specific neurons act is reduced. Therefore, a Dropout layer is added in each layer of the LSTM network, and the Dropout layer is used for controlling the node weight of the hidden layer, so that certain track characteristics are prevented from being effective only under a fixed combination, and the network is consciously led to learn the common commonality of the tracks. And converting the dimension into the required output latitude after the characteristic quantity output to the LSTM passes through a full connection layer.

After the network structure and input and output are determined, the network can be trained by using a training sample set, and a mean square error function shown as a formula (3) is selected as a loss function:

wherein N represents the number of batch samples in one training process, Y _pred Predicted value, Y, representing neural network output _i Representing the corresponding real value, the neural network training is the process of updating the weight, and the optimization goal of the network is to make E approach to 0. When network training is carried out, an adaptive moment estimation method (Adam) is selected ^[13] The method comprises the following steps:

firstly, considering that in the traditional back propagation algorithm, the updating direction of the weight only depends on the gradient obtained by the current sample, so the concept of momentum (moment) in physics is used for reference, namely, the direction updated before is kept to a certain extent when the weight is updated, and the final updating direction is obtained by adding the gradient of the current sample, namely, the final updating direction is obtained

Where s is called momentum, and is also an estimate of the first moment of the gradient, β ₁ Called first order momentum decay coefficient, typically taken as beta ₁ ＝0.9。

Where v is called the velocity quantity, and is also an estimate of the second moment of the gradient, β ₂ Called second order momentum decay coefficient, typically taken as beta ₂ ＝0.999。

Further, to realize adaptive adjustment of the learning rate, i.e. to use a larger epsilon update for the weight with a slower update and a smaller epsilon update for the weight with a faster update, the learning rate epsilon is adjusted:

δ is a constant to prevent the denominator from being zero, and δ is generally taken to be 10 ^-8 。

On the basis of this, the first and second moment estimates of the gradient are corrected unbiased:

and finally, obtaining a weight updating formula based on the Adam algorithm:

according to the steps, the training of the LSTM network is completed, the trained model is stored locally, some parameters such as sample mean values, variances and the like are also stored in a txt file format, and when a new sample to be tested is tested, the local files can be directly called to quickly complete a simulation experiment.

The values of the network structure parameters during model training and testing are shown in table 3. And inputting new data to be tested for the trained prediction model, and verifying the model prediction accuracy.

TABLE 3 LSTM network parameters

When a sample data set is selected, in a simulation countermeasure environment, a group of experimental data, namely a track curve, can be obtained every time a simulation countermeasure is completed, and the fact that the characteristic points of a single curve are too few and are not beneficial to learning of a neural network is considered, so that multiple experiments are carried out, and tracks are spliced. Time-series data including 10600 points each including twenty-three-dimensional feature quantities was finally obtained. After dimensionality reduction, a sample set with dimensions (21200,23) is obtained. The training set and test set were partitioned in a ratio of 0.75:0.25 to yield a test set sample dimension of (5300, 23).

In order to prevent the model from being over-fitted in the set epoch range, in addition to adding a Dropout layer when designing the model, sample data is processed. For training data, a part of the verification set data is selected as verification data (valid data), the proportion of the verification set data to the training set data is defined as valid _ data _ rate ═ 0.15, a patience value ═ 5 is set, and if the iteration number of the verification set exceeds 5 times and the training precision and loss are not improved, the training is stopped in advance, which is called 'Early training Stopping' (Early training Stopping) ^[14] By adding the verification set, the overfitting condition can be avoided, and model selection is perfected. For the training set and the validation set, as shown in the left diagram of fig. 4, the loss condition under epoch is given, the red line is the training set loss, and the blue line is the validation set loss. To analyze the model's loss on the training set individually, the previous 1000 training loss values are saved for each batch training, as shown in the right panel of FIG. 4. It can be seen that as the number of times of model training increases, the model loss value gradually decreases, approaching to 0, which indicates that the model training effect is ideal.

Fig. 5 shows the predicted results of the aircraft in the longitude, latitude and altitude directions, wherein the red solid line is the actual aircraft trajectory, i.e., the label value, and the blue dotted line is the aircraft trajectory predicted by the LSTM model. The jump occurs at the 2800 th time in the figure, because the sample track is formed by splicing small tracks of a segment, and therefore, the jump occurs at some time. As can be seen from the simulation result, the difference between the predicted value and the true value is small, and the model prediction precision is high.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Reference documents

[1]Han J,Pei J,Yin Y.Mining frequent patterns without candidate generation[J].ACM sigmod record,2000,29(2):1-12.

[2] Study and improvement of Apriori algorithm in association rules mining [ J ] computer application, 2010,30(11): 2952-.

[3]Morzy M.Mining frequent trajectories of moving objects for location prediction[C]. International workshop on machine learning and data mining in pattern recognition.Springer, Berlin,Heidelberg,2007:667-680.

[4]Yang J,Xu J,Xu M,et al.Predicting next location using a variable order Markov model[C].Proceedings of the 5th ACM Sigspatial International Workshop on GeoStreaming. ACM,2014:37-42.

[5]Qiao S J,Shen D,Wang X,et al.A Self-Adaptive Parameter Selection Trajectory Prediction Approach via Hidden Markov Models[J].Intelligent Transportation Systems,IEEE Transactions on,2015,16(1):284-296.

[6]Payeur P,Le-Huy H,Gosselin C M.Trajectory prediction for moving objects using artificial neural networks[J].IEEE Transactions on Industrial Electronics,1995,42(2): 147-158.

[7]Huang W,Song G,Hong H,et al.Deep architecture for traffic flow prediction:deep belief networks with multitask learning[J].IEEE Transactions on Intelligent Transportation Systems,2014,15(5):2191-2201.

[8] Billows, guokui, xusimon, et al. fighter aircraft airfight trajectory prediction based on Elman neural network [ J ] flight mechanics, 2018,36(5):89-94.

[9]Wang J C,Qi X H,Shan G.Elman network based on EPSO-BP and its application in flight trajectory prediction[J].Control and Decision,2013,28(12):1884-1888.

[10] Kui, Zhou Ying, Yangshizhi, et al. A BP neural network-based air target track prediction model [ J ] command information system and technology, 2017,8(3):54-58.

[11] Yangyongnong, Yuanlongfei, Songmin, etc. an unmanned aerial vehicle track prediction model and simulation [ J ] based on Bi-LSTM, the progress of aeronautical engineering, 2020,11(01):77-84.

[12]Li X R,Jilkov V P.Survey of maneuvering target tracking.Part I.Dynamic models[J].IEEE Transactions on aerospace and electronic systems,2003,39(4):1333-1364.

[13]Kingma D P,Ba J.Adam:A method for stochastic optimization[J].arXiv preprint arXiv:1412.6980,2014.

[14]Yao Y,Rosasco L,Caponnetto A.On early stopping in gradient descent learning[J]. Constructive Approximation,2007,26(2):289-315。

Claims (2)

1. A method for predicting the track of an aircraft based on a long-term and short-term memory network is characterized in that noise interference carried by a sensor feature vector is eliminated by Kalman filtering; for the directly acquired state parameters, data preprocessing is carried out, including down-sampling, invalid value elimination and missing value complementation, in addition, in order to improve the calculation stability, the data is subjected to normalization processing, and the value range of the input data is included in a [0,1] interval; constructing a locus prediction network model based on LSTM, defining the input and output of the network model, and carrying out supervision training on the network model; the detailed steps are as follows:

firstly, acquiring required characteristic data from a hostile and antagonistic simulation platform of an enemy aircraft, and summarizing the aircraft track obtained by considering the influence of various factors such as the aircraft track, battlefield environment and enemy motivation to have the three characteristics of continuity, time sequence and interactivity:

continuity means that the unmanned aerial vehicle track is continuously changed, but not discontinuous;

the time sequence is that the track data is time-nature, and the position of the next moment is related to the position of the previous moment, so the track data is a time sequence data in nature;

interactivity means a complex process of dynamic change among multiple airplanes in an actual environment, the maneuver of one airplane can influence the maneuver of another airplane, and the position of one airplane can also influence the position of another airplane;

therefore, when the trajectory prediction is performed, in addition to the position, the posture and the speed of the enemy plane, the 23-dimensional feature vectors of the two planes are considered due to the interactivity:

longitude, latitude, altitude, pitch angle, roll angle, course angle, azimuth angular velocity, pitch angular velocity, north directional velocity, sky directional velocity, east directional velocity, north directional acceleration, sky directional acceleration, east directional acceleration, horizontal entrance angle, relative distance change rate, relative altitude, radar state identification result, radial velocity, true course, and ground speed;

secondly, after the feature vector is obtained, since the raw data obtained from the sensor cannot be directly used as network input, it needs to be preprocessed, according to 1: 5, carrying out data downsampling on data at sampling intervals, removing invalid values in the data, and complementing missing values by using a mean filling method, in addition, because the numerical values of part of used parameters are large, in order to improve the calculation stability, carrying out normalization processing on the data, and enabling the value range of input data to be included in a [0,1] interval, wherein the normalization formula is as follows:

wherein X is the actual value of a certain characteristic quantity, and X is _max ,X _min Respectively the maximum value and the minimum value of X in all data, and Y is the result after normalization;

and (3) performing inverse normalization on the prediction result by using the trained prediction model so as to perform error comparison analysis with the actual value, wherein the inverse normalization formula is as follows:

X＝(X _max -X _min )Y+X _min (2)

after the data processing is completed, a normalized sample data set available for the neural network is obtained, at this time, input and output of the LSTM network need to be constructed according to actual requirements, when label data are selected, longitude, latitude and height of the next moment are taken as feature labels of the current moment, segmentation of the data set is completed by adopting a sliding window method, input of the network is obtained by iterative selection of original data, the time _ step is from 0 to the first time _ step, the time _ step selected by the training is 20, namely from 0 to 19 is the first input, from 1 to 20 is the second input, and the like; the dimensionality of input data is (N,20,23), wherein N depends on the size of a sample set, 20 refers to a time _ step value of an LSTM, 23 refers to the dimensionality of a feature vector, the label data is needed by a training network, and the object of the model is to predict the future trajectory of the airplane one step, so that the latitude and longitude heights of an airplane carrier are selected as labels at the input of the first 0-19, the longitude and latitude heights of the airplane carrier are selected as labels at the input of the second 1-20, and the like, so that the dimensionality of output data is (N,20,3), wherein N depends on the size of the sample set, 20 refers to the time _ step value of the LSTM, and 3 corresponds to the output dimensionality, namely the longitude, the latitude and the height of the airplane carrier;

and finally, constructing an LSTM network, using the LSTM as a main body part of the model, learning the relation between the input track related characteristic quantity and the label of the position at the next moment by constructing a multi-layer network structure, predicting the position which is possibly reached in the future, wherein the prediction model mainly comprises an LSTM layer, a Dropout layer and a full connection layer, and selecting two layers of LSTMs for model building: the LSTM network of the first layer takes twenty-three-dimensional characteristic quantity asInput, output of the second layer and the first layer

As input, second layer output

After passing through a full connection layer, obtaining a final network output Y, adding a Dropout layer in each LSTM network, controlling the node weight of a hidden layer by using the Dropout layer, avoiding certain track characteristics from being effective only under fixed combination, consciously enabling the network to learn the universality of tracks, and converting the dimensionality into the required output latitude after passing through the full connection layer on the characteristic quantity output by the LSTM;

after the network structure and input and output are determined, training the network by using a training sample set, and selecting a mean square error function shown as a formula (3) as a loss function:

wherein N represents the number of batch samples in one training process, Y _pred Predicted values representing neural network outputs, Y _i Representing the corresponding real value, the neural network training is the process of updating the weight, the optimization goal of the network is to make E approach to 0, when the network training is carried out, an adaptive moment estimation method (Adam) is selected, and the adaptive moment estimation method comprises the following steps:

firstly, considering that in the traditional back propagation algorithm, the updating direction of the weight only depends on the gradient obtained by the current sample, so the concept of momentum (moment) in physics is used for reference, namely, the direction updated before is kept to a certain extent when the weight is updated, and the final updating direction is obtained by adding the gradient of the current sample, namely, the final updating direction is obtained

Where s is called momentum, and is also an estimate of the first moment of the gradient, β ₁ Referred to as the first order momentum decay coefficient;

where v is called the velocity quantity, and is also an estimate of the second moment of the gradient, β ₂ Called second order momentum decay coefficient, typically taken as beta ₂ ＝0.999；

Further, to realize adaptive adjustment of the learning rate, i.e. to use a larger epsilon update for the weight with a slower update and a smaller epsilon update for the weight with a faster update, the learning rate epsilon is adjusted:

δ is a constant for preventing the denominator from being zero, and is generally taken to be 10 ^-8 ；

On the basis of the above, the first and second moment estimates of the gradient are corrected unbiased:

and finally, obtaining a weight updating formula based on the Adam algorithm:

according to the steps, the training of the LSTM network is completed, the trained model is stored locally, the sample mean value and variance parameters are stored in a txt file format, and the local files are directly called when a new sample to be tested is tested;

when model training and testing are performed, the values of network structure parameters are shown in the following table:

LSTM network parameter table

Inputting new data to be tested for the trained prediction model, and verifying the model prediction accuracy;

when a sample data set is selected, in a simulation countermeasure environment, each time a simulation countermeasure is completed, a group of experimental data, namely a track curve, is obtained, the condition that characteristic points of a single curve are too few and are not beneficial to learning of a neural network is considered, therefore, multiple experiments are carried out, tracks are spliced, time sequence data containing 10600 points are finally obtained, each point contains twenty-three-dimensional characteristic quantity, after dimensionality reduction, a sample set with dimensionality (21200,23) is obtained, a training set and a testing set are divided according to the proportion of 0.75:0.25, and the sample dimensionality of the testing set is (5300, 23);

in order to prevent the model from being over-fitted within the set epoch range, in addition to adding a Dropout layer when designing the model, sample data is processed, for training data, a part of the training data is selected as verification data (valid data), the proportion of the verification set data to the training set data is defined as valid _ data _ rate being 0.15, a endurance value probability being 5 is set, if the iteration number of the verification set exceeds 5 times and the training precision and loss are not improved, the training is stopped in advance, which is called as Early training Stopping (Early Stopping), and by adding the verification set, the over-fitting condition can be avoided, and the model selection is completed.

2. The method for predicting aircraft trajectories based on the long-short term memory network as claimed in claim 1, wherein the method comprises the following steps:

the method comprises the steps that firstly, original state information is obtained from an on-board sensor, twenty-three-dimensional characteristic quantities are selected as the input of a prediction model;

the second step is to preprocess the data: firstly, the original sensor parameters contain certain noise interference, and a Kalman filtering method is selected to eliminate the noise interference; secondly, in order to improve the data utilization rate and reduce the calculation cost, the data processing method comprises the following steps of 1: 5, performing down-sampling, and performing normalization processing on the data; finally, rejecting the invalid value, complementing the missing value and finishing the data preprocessing work; then, input and tag value selection are carried out: the segmentation of a data set is completed by adopting a sliding window method, the input of a network is obtained by iterative selection of original data, from 0 to the first time _ step, the time _ step selected in the training is 20, namely, from 0 to 19 is the first input, from 1 to 20 is the second input, and so on; the dimensionality of input data is (N,20,23), wherein N depends on the size of a sample set, 20 refers to a time _ step value of an LSTM, 23 is the dimensionality of a feature vector, label data are needed by a training network, and the purpose of the model is to predict the track of the airplane in one step in the future, so that the longitude and latitude heights of the airplane carrier are selected to be labels from 1 to 20 for the first input from 0 to 19, the longitude and latitude heights of the airplane carrier are selected to be labels from 2 to 21 for the second input from 1 to 20, and the like;

and thirdly, constructing a track prediction model based on the LSTM, wherein the constructed model comprises two LSTM layers, a Dropout layer is added behind each LSTM layer, the phenomenon of overfitting is avoided, and the dimension is converted into the required output dimension after the output characteristic quantity passes through a full connection layer.

Images