Method for improving track identification effect through preprocessing and data enhancement in maneuvering rule identification

Abstract

According to the method, based on a mode of combining maneuvering trajectory data enhancement simulation and deep learning, characteristics of several typical maneuvering trajectories are analyzed, and data enhancement is mainly carried out by carrying out changes such as spatial rotation and translation on data; the identification effect of a target motion law mode is enhanced through uniformization operation and translation invariant preprocessing operation on track data, and track features are extracted based on a scattering transformation and long-short-term memory network method, so that maneuvering mode identification is carried out. And trajectory simulation is carried out on the trajectory of flexible parameter model modeling, so that the flexibility of maneuvering rule identification can be greatly improved. Simulation shows that the provided preprocessing method plays an important role in the identification effect of the trajectory maneuvering rules with various forms and changes, and the accuracy rate reaches 95%.

Description

Technical Field

[0001] The present invention relates to a method which can be used to improve the effect of maneuvering law identification trajectory. Background Art

[0002] Common maneuvers include straight line, left turn, right turn, up turn, down turn, left detour, right detour, J maneuver, S maneuver, half-reverse, and roller maneuver.

[0003] Identifying the motion patterns of maneuvering enemy targets is crucial for determining their threat level. Determining the enemy's maneuvering patterns based on their motion patterns, and then assessing their threat level based on this information, plays a vital role in air combat and has a positive impact on national defense and military defense.

[0004] There are many methods involved in determining the maneuverability of target trajectories, such as dynamic time warping algorithm, segmented trajectory recognition, segmented maneuver recognition algorithm, hypothesis generation and elimination algorithm, etc.

[0005] A maneuver sequence of a target trajectory is extracted as a test template and matched with a reference template using a dynamic time warping algorithm. This allows the target's maneuver type to be identified. Comparisons with a hidden Markov model show that the dynamic time warping algorithm achieves higher recognition rates and shorter recognition times, with the same rejection rate. Segmented trajectory recognition divides the trajectory into several curved segments representing different motion models. The curve parameters are estimated using least squares to obtain target parameter estimates, but this approach is only applicable to uniform motion and uniform turning states and is not suitable for tracking accelerating targets. The segmented maneuver recognition algorithm builds on the segmented trajectory recognition algorithm by incorporating a curved model for accelerating targets. By incorporating velocity continuity into the curved segments, the cost function is improved, the segmentation end judgment criteria are refined, and the motion parameters are accurately estimated, resulting in superior tracking performance compared to segmented trajectory recognition. A hypothesis generation and elimination algorithm can address the segmentation problem, encoding the maneuver types of aerial targets. Based on this, a model for calculating the similarity between maneuvers and maneuver types is proposed.

[0006] The method of deep learning is used to identify the maneuver type, and the maneuver mode is identified based on scattering transform and long-short-term neural network, which has a good effect on identifying the movement patterns of enemy maneuvering targets.

[0007] However, this method still does not take into account the situation where the data is not within the center circle, the length is inconsistent, and the shape of the trajectory is not completely consistent due to noise and inconsistent motion radius. Summary of the Invention

[0008] The technical problem solved by the present invention is: to facilitate the classification of trajectories of different directions, different motion parameters, and different lengths, a method for improving the identification effect of trajectories by preprocessing and data enhancement in maneuvering pattern identification is provided. Trajectory simulation is performed on trajectories modeled by flexible parameter models, which greatly increases the flexibility of maneuvering pattern identification.

[0009] The technical solution of the present invention is characterized in that it comprises the following steps:

[0010] Training data set generation and preprocessing operations on training data set:

[0011] 1) For the trajectory basic data set to be identified, one trajectory is generated for each trajectory, and the following enhancement operations are performed on each trajectory;

[0012] 2) By expanding the direction of the basic data set, adding noise, translation, and increasing radius changes, and cross-forming a new enhanced trajectory data set, the effect and accuracy of identifying the motion law of maneuvering targets can be improved;

[0013] 3) Perform translation-invariant preprocessing on the training data set, so that the center of the trajectory is at the origin of the coordinate system through translation operation;

[0014] 4) Perform a normalization operation on the trajectory data length of the training dataset to ensure the smoothness and rationality of the data. Finally, after preprocessing, N (3, L) trajectory data are obtained, where N is the number of trajectories and L is the normalized trajectory length.

[0015] The data to be identified should be pre-processed before testing the test dataset, including:

[0016] 5) Perform translation-invariant preprocessing on the test data set, so that the center of the trajectory is at the origin of the coordinate system through translation operation;

[0017] 6) Perform a normalization operation on the trajectory data length of the test dataset to ensure the smoothness and rationality of the data. Finally, after preprocessing, N (3, L) trajectory data are obtained, where N is the number of trajectories and L is the normalized trajectory length.

[0018] Different category labels need to be added to the training and test datasets. Data augmentation is performed on the training set. For trajectories with different radii and noise levels, pairs of trajectories and corresponding labels are provided to the training network to adapt to trajectories with different radii and noise levels.

[0019] In step 2), the training data set is pre-processed with different noise and radius data enhancements: the radius is changed to about 50% and 30% of the original turning radius; in order to better add noise according to the actual situation, it is necessary to preliminarily estimate the variance of the relationship between the measurement noise and the aircraft speed, and add Gaussian white noise with two variances;

[0020] In step 3) and step 5), the translation-invariant preprocessing operation of the trajectory data is to perform translation-invariant preprocessing on the training data and the test data before inputting the network. The method of the translation-invariant preprocessing operation is as follows:

[0021]

[0022] The coordinates of the position after centering are marked as (x′, y′, z′), and the coordinates before centering are (x, y, z), where the subscript l represents the sequence number of each discrete moment of the trajectory;

[0023] The trajectory data length unification operations in steps 4 and 6) are performed by unifying the three coordinate dimensions of the coordinate system in the time dimension without changing the changing characteristics of the input trajectory. That is, regardless of the time and sampling interval of the data, it is unified into data with a time dimension of L. L is a specific parameter, and an appropriate length that can reflect the regular characteristics of various trajectories is selected. Specifically, the relationship between the original length L' of the data to be unified and the length L to be unified is considered. If L' / L>1, compressed sampling is performed; if L' / L<1, the original trajectory data is repeatedly expanded. In this process, the smoothness and rationality of the data are guaranteed.

[0024] The principle of the present invention is that during the process of identifying the motion patterns of enemy maneuvering targets, data may not be within the central circle, length may be inconsistent, and the trajectory shape may not be completely consistent due to noise or inconsistent motion radius. Based on this, the proposed translation invariance preprocessing, length uniformity preprocessing, data enhancement preprocessing with different radii, and data enhancement preprocessing with different noise levels can enhance the flexibility of training data in terms of different noise levels and radii by analyzing data characteristics. Furthermore, training based on scattering changes and long-short-term neural networks can improve the effectiveness of the recognition method and enhance the recognition effect.

[0025] Translation-invariant preprocessing performs preprocessing operations on data position centering data, and de-centers the training data and test data before inputting the network. Length consistency preprocessing can make the length of the input trajectory data consistent, and ensure the smoothness and rationality of the data in the input scattering feature extraction network and LSTM network. By performing data enhancement preprocessing with different radii and different noise data, the effect and accuracy of identifying the motion laws of maneuvering targets can be improved.

[0026] The advantages of the present invention over the prior art are: the use of deep learning to identify maneuver types, identification of maneuver modes based on scattering transform and long-short-term neural networks, and good results in identifying the movement patterns of enemy maneuvering targets. The present invention takes into account situations where the data is not within the center circle and the length is inconsistent, as well as situations where the shape of the trajectory is not completely consistent due to noise and inconsistent motion radius, and performs trajectory simulation on the trajectory modeled by the flexible parameter model, which greatly increases the flexibility of maneuver pattern identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of a method for improving the identification trajectory effect by preprocessing and data enhancement in the identification of maneuver patterns of the present invention.

[0028] Figure 2 The identification results of the maneuvering law of Experiment 1 under the ideal trajectory conditions of the present invention are

[0029] Figure 3 The identification results of the maneuvering rules of Experiment 2 under the data enhancement conditions of the present invention are as follows:

[0030] Figure 4 The fine-grained recognition results of the maneuvering rules of Experiment 3 under the data enhancement conditions of this invention are shown in Figure 2.

[0031] Figure 5 This is the confusion of the large category identification of the maneuvering rules in Experiment 3 under the data enhancement conditions of the present invention.

[0032] Figure 6 This is the result of identifying the general category of the maneuvering law of the experimental 3 data under the condition that the data of the present invention is not enhanced. DETAILED DESCRIPTION

[0033] Consider situations where the input training data is not within the center circle, has inconsistent lengths, and has inconsistent trajectory shapes due to noise or inconsistent motion radii. Data preprocessing is required to ensure translation invariance, length consistency, and enhance data with varying radii and noise levels.

[0034] The following are the specific steps for data preprocessing:

[0035] Training data set generation and preprocessing operations on training data set:

[0036] 1) For the trajectory basic data set to be identified, one trajectory is generated for each trajectory, and the following enhancement operations are performed on each trajectory;

[0037] 2) By expanding the basic dataset, adding noise, translation, and radius variations to generate a new, enhanced trajectory dataset, the effectiveness and accuracy of identifying maneuvering target motion patterns can be improved. Data augmentation preprocessing, including adding noise and radius variations, can also improve the accuracy of identifying maneuvering target motion patterns. Necessary spatial transformations, including proportional changes in turning radius, were performed to ensure consistency with actual conditions and simplify data generation complexity.

[0038] The training dataset was preprocessed with different noise and radius data augmentation: the radius was varied to approximately 50% and 30% of the original turning radius, and the radius was varied to approximately 50% and 30% of the original turning radius. To better adapt the noise addition to the actual situation, a preliminary estimate of the variance of the relationship between the measurement noise and the aircraft speed was performed, and Gaussian white noise with two variances was added.

[0039] Based on the training and test datasets, different category labels need to be added for data enhancement. For trajectories with different radii and different noises, the program provides pairs of trajectories and corresponding labels to the training network for training, so as to achieve the purpose of training the network to adapt to trajectories with different radii and noises.

[0040] 3) Perform translation-invariant preprocessing on the training data set, so that the center of the trajectory is at the origin of the coordinate system. The translation-invariant preprocessing operation on the trajectory data is performed on the data position centering preprocessing operation. The training data and the test data are de-centered before entering the network, and several category labels are added to the training data and the test data respectively.

[0041] The preprocessing method is

[0042]

[0043] The coordinates of the position after centering are marked as (x′, y′, z′), and the coordinates before centering are (x, y, z), where the subscript l represents the sequence number of each discrete moment of the trajectory;

[0044] 4) Perform a trajectory data length unification operation on the training dataset, ensuring the smoothness and rationality of the data during this process. Finally, after preprocessing, N (3, L) trajectory data are obtained; N is the number of trajectories, and L is the normalized trajectory length. The trajectory data length unification operation unifies the time length dimension of the three coordinate dimensions of the Northeast and Sky without changing the variation characteristics of the input trajectory. That is, regardless of the amount of time and sampling interval, the data is unified into data with a time dimension of L; L is a fixed parameter, and an appropriate length that can reflect the regular characteristics of various trajectories is selected. Specifically, the relationship between the original length L' of the data to be unified and the length L to be unified is considered. If L' / L>1, compressed sampling is performed; if L' / L<1, the original trajectory data is repeatedly expanded. In this process, the smoothness and rationality of the data are guaranteed.

[0045] The data to be identified should be pre-processed before testing the test dataset, including:

[0046] 5) Perform translation-invariant preprocessing on the test data set, so that the center of the trajectory is at the origin of the coordinate system through translation operation. The specific steps are the same as the preprocessing operation on the training data;

[0047] 6) Perform a normalization operation on the trajectory data length of the test dataset to ensure the smoothness and rationality of the data. Finally, after preprocessing, N (3, L) trajectory data are obtained, where N is the number of trajectories and L is the normalized trajectory length. The specific steps are the same as the preprocessing operation for the training data.

[0048] In order to illustrate the advantage of the present invention in improving the flexibility of maneuver law identification, an experimental description is given.

[0049] The dataset was generated using our own simulations. For each maneuver, based on the example trajectory, each trajectory was expanded to 19 directions. Gaussian white noise of two different intensities was then added. For maneuvers with potentially large variations in turning radius, two additional turning radius scenarios and two translation scales were added, creating four new trajectory data sets. The dataset contains 15 maneuvers, so the initial dataset is quite large.

[0050] The network structure and main parameters in this experiment are shown in Table 1. After feature extraction, a multi-dimensional tensor is formed. The trajectory length used is L = 45.

[0051] Table 1 Network structure and parameters

[0052] Number of levels Network structure Network parameters Parameter value one LSTM input layer Input Tensor Features*L two LSTM hidden layer Hidden layer width 100 three LSTM output layer MaxEpochs 500～600 Four FC layer BatchSize 1000 five Softmax layer Initial learning rate 0.01 six Judgment result output Optimized algorithm Adam

[0053] Experiment 1 designed the training data labels as J-type (Jxing), J-type reverse (Jxingfanxiang), S-type (S), S-type reverse (Ss), half-reversal (bankou), half-reversal reverse (bankouzxf), straight line (directline), downturn (downturn), roller maneuver (guntong), roller reverse (guntongfanxiang), left turn (leftturn), left detour (leftyihui), right turn (rightturn), right detour (rightyuhui), upturn (upturn). When no noise is added to the trajectory and the original trajectory radius is fixed, after the original running direction is enhanced at various horizontal angles, the test accuracy is as follows: Figure 2 . Figure 2 The double columns on the right indicate whether other categories were predicted.

[0054] Since the roller reverse (guntongfanxiang) and the roller (guntong) belong to the same roller type, the recognition accuracy of the maneuvering law in this case can be determined to be 100%.

[0055] Experiment 2 performs data enhancement based on the data of Experiment 1:

[0056] (1) Several cases of radius change and noise addition are added to the data set. The radius changes are about 50% and 30% of the original turning radius, and are marked as r1 and r2 in the order of radius change ratio from small to large.

[0057] (2) Noise is marked as follows: from weak to strong;

[0058] (3) Translation distance enhancement adds two cases: adding 30 to each of the xyz coordinates and adding 100 to each of the three coordinates to the original trajectory set, marked as py1 and py2.

[0059] Then, on the expanded dataset, on the training data, and on the original basic version type, we added several labels including “Ss_n1_r”, “leftyihui_n1_r1”, and “leftyihui_n1_r2”.

[0060] Based on the training data categories, the test data adds several category labels: "Ss_n1_r1", "Ss_n1_r2", "Ss_n2_r1", "Ss_n2_r2", "left_yihui_n2_r1", "left_yihui_n2_r2", "leftyihui_n1_r1_py1", "leftyihui_n1_r1_py2", "leftyihui_n1_r2_py1", "leftyihui_n1_r2_py2", "left_yihui_n2_r1_py1", "left_yihui_n2_r1_py2", "left_yihui_n2_r2_py1", and "left_yihui_n2_r2_py2".

[0061] There are 13 types of enhanced extensions in the left detour maneuver type (leftyihui, leftyihui_r2_n2, leftyihui_r2_n2_py1, leftyihui_r2_n2_py2, ...). After enhanced training, almost 100% of them can be correctly identified as leftyihui. Figure 3 As shown in the experimental results, from the recognition of S maneuver type, it can be seen that without any augmented training on radius changes, the cases of S_n1_r1 are all classified as S type. Figure 3 The double columns on the right indicate whether other classes were predicted. The experiments demonstrate that the designed preprocessing method can improve recognition performance. This allows the designed method to essentially solve the problem of maneuver recognition in a variety of situations, achieving the expected and largely satisfactory recognition accuracy.

[0062] Experiment 3 continued data augmentation based on the reasoning of Experiment 2. The training set used several combinations of radius n1, r1, r2, and py1 described in Experiment 2 to provide a relatively comprehensive coverage of the number of categories, with the exception of Jxing, Jxingfanxiang, guntong, guntongfanxiang, and directline. Due to space limitations, we will not list them one by one. The detailed label categories of the test set are the same as those of the training set. There are a total of 65 specific label situations. After one test, the test results are as follows Figure 4 shown. Figure 5 This is the result of a broad category judgment, ignoring the different steering, radius, and noise suffixes in the detailed labels. The judgment is correct only if the broad category of maneuver is correct. Figure 5 The various misjudgments are clearly displayed. After statistics, it was found that the number of misclassified samples was 11, the total number of samples was 203, the error rate of major classification was 0.05, and the recognition accuracy reached 95%.

[0063] In order to compare the effectiveness of the proposed method, the basic network trained with the original unenhanced dataset, that is, the basic training of the maneuver law recognition experiment 1, was used to perform label prediction tests on the most comprehensive enhanced data label dataset obtained in experiment 3. The results of the major classifications are shown in the figure below. Figure 6 As shown in Figure 2. Note that the most comprehensive augmented data set here is uncentered. We found a high error rate for large-category recognition, with an accuracy rate of only 62.5%. This contrasts sharply with the improved method, demonstrating that our improved method significantly improves its applicability to various parameter trajectories.

Claims

1. A method for improving the recognition trajectory effect by preprocessing and data enhancement in maneuver law identification, characterized in that The following steps are involved: Training data set generation and preprocessing operations on training data set: 1) For the trajectory basic data set to be identified, one trajectory is generated for each trajectory, and the following enhancement operations are performed on each trajectory; 2) By expanding the direction of the basic data set, adding noise, translation, and increasing radius changes, and cross-forming a new enhanced trajectory data set, the effect and accuracy of identifying the motion law of maneuvering targets can be improved; 3) Perform translation-invariant preprocessing on the training data set, so that the center of the trajectory is at the origin of the coordinate system through translation operation; 4) Perform a uniform operation on the trajectory data length of the training dataset to ensure the smoothness and rationality of the data. Finally, after preprocessing, N (3, L) trajectory data are obtained. N is the number of trajectories, L is the normalized trajectory length; The data to be identified should be pre-processed before testing the test dataset, including: 5) Perform translation-invariant preprocessing on the test data set, so that the center of the trajectory is at the origin of the coordinate system through translation operation; 6) Perform a normalization operation on the trajectory data length of the test dataset to ensure the smoothness and rationality of the data. Finally, after preprocessing, N (3, L) trajectory data are obtained, where N is the number of trajectories and L is the normalized trajectory length.

2. The method for improving trajectory identification performance by preprocessing and data enhancement in maneuver pattern identification according to claim 1, characterized in that: Based on the training dataset and the test dataset, different category labels need to be added; data enhancement is performed on the training set, and for trajectories with different radii, different translations, and different noises, trajectories and corresponding labels are provided in pairs to the training network for training, so as to achieve the purpose of training the network to adapt to trajectories with different radii and noises.

3. The method for improving trajectory identification performance by preprocessing and data enhancement in maneuver pattern identification according to claim 1, characterized in that: In step 2), the training dataset is preprocessed with different noise and radius data enhancements: the radius is changed to about 50% and 30% of the original turning radius; to better add noise based on the actual situation, it is necessary to preliminarily estimate the variance of the relationship between the measurement noise and the aircraft speed, and add Gaussian white noise with two variances.

4. The method for improving trajectory identification performance by preprocessing and data enhancement in maneuver pattern identification according to claim 1, characterized in that: In step 3) and step 5), the translation-invariant preprocessing operation of the trajectory data is to perform translation-invariant preprocessing on the training data and the test data before inputting the network. The method of the translation-invariant preprocessing operation is as follows: The coordinates of the position after centering are marked as (x′, y′, z′), and the coordinates before centering are (x, y, z), where the subscript l represents the sequence number of each discrete moment of the trajectory.

5. The method for improving trajectory identification performance by preprocessing and data enhancement in maneuver pattern identification according to claim 1, characterized in that: The trajectory data length unification operation in steps 4 and 6) is to unify the three coordinate dimensions of the coordinate system data in the time dimension without changing the change characteristics of the input trajectory. That is, regardless of the time and sampling interval of the data, it is unified into data with a time dimension of L. L is a fixed parameter, and an appropriate length that can reflect the regular characteristics of various trajectories is selected. Specifically, the relationship between the original length L' of the data to be unified and the length to be unified L is considered. If L' / L>1, compressed sampling is performed. If L' / L<1, the original trajectory data is repeatedly expanded; in this process, the smoothness and rationality of the data are ensured.

Images