Unmanned aerial vehicle target trajectory identification method and system based on RoPE-BiLSTM model

Abstract

The invention provides an unmanned aerial vehicle target trajectory identification method and system based on an RoPE-BiLSTM model, based on trajectory data of a plurality of unmanned aerial vehicle targets and non-unmanned aerial vehicle targets, trajectory dynamic time sequence features and trajectory static statistical features are respectively extracted by adopting a specific calculation method, and the RoPE-BiLSTM model and an improved multi-layer perceptron model are constructed. Dynamic behavior information is extracted by inputting the dynamic features into a RoPE-BiLSTM model; the method comprises the following steps: performing feature extraction on static statistical features through an improved multilayer perceptron model, performing feature fusion on results output by two models through a Bayesian linear layer, and separating relative displacement and absolute position between differential sampling track points to avoid information loss of data in a normalization process. Therefore, the model can accurately identify the unmanned aerial vehicle trajectory.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle target recognition, and in particular to a method and system for unmanned aerial vehicle target trajectory recognition based on a RoPE-BiLSTM model. Background Art

[0002] With the rapid development of drone technology, low-altitude drones are increasingly used in civil and military fields. Their flexibility and efficiency make them an indispensable tool in modern combat and daily life. At the same time, the challenge of how to effectively detect and identify these low-altitude aircraft has become increasingly prominent. Current anti-drone technologies mainly include electromagnetic interference, physical interception, and intelligent monitoring. However, existing methods still face many limitations. Many traditional technologies rely on static monitoring systems and are easily affected by environmental factors, resulting in reduced recognition accuracy. Secondly, in complex combat environments, existing methods often fail to effectively distinguish when dealing with diverse flight trajectories and different types of drones, resulting in false alarms and missed reports. In addition, the detection capabilities of low-altitude flights and small drones are generally insufficient, which limits the practical application effect of existing technologies.

[0003] In terms of hardware, radar is the main means of drone detection, which can provide rich trajectory information for drone detection, such as slant range, relative altitude, radial velocity and radar cross section (RCS). These time series information contain rich drone behavior characteristics, laying the foundation for drone target trajectory recognition. At present, the mainstream time series detection methods based on this information are based on time series neural networks, such as LSTM, Transformer, etc. However, these methods lack sufficient robustness and stability when facing edge data; at the same time, due to their black box characteristics, there is not enough physical support, it is difficult to make further constraints, and there is a lack of effective countermeasures when the model performance drops dramatically.

[0004] For time series data, positional encoding has gradually shown its importance in practical applications. Positional encoding has many different implementations in subsequent articles based on the Transformer architecture. Especially now that large language models are popular, choosing appropriate positional encoding will also improve the training effect when facing long token input. Absolute Positional Encoding is the most common positional encoding method. The idea is to add a position vector to each element of the input sequence to indicate the specific position of the element in the sequence. The position vector is usually generated by a fixed function and is independent of the input data. Sine and cosine functions are usually used, so that the generated encoding has strong periodicity and can capture the relative position information in the sequence. Rotary Positional Encoding (RoPE) is a new positional encoding method proposed in recent years. It embeds the position information into the vector by rotating the input vector. Specifically, RoPE encodes the position information in the sequence by rotating the vector in each dimension.

[0005] At present, the research technology of drone target trajectory recognition is still at the stage of simple time series or static neural network models, lacking sufficient generalization ability and interpretability. 1) There is a lack of time series neural network classification models with high recognition accuracy and high generalization for different scenarios; 2) There is a lack of model interpretability foundation based on the physical meaning of the real world; 3) Models based on large amounts of time series data are not suitable for model deployment scenarios with limited computing and storage conditions. Summary of the invention

[0006] The present invention aims at the problems existing in the current process of unmanned aerial vehicle target trajectory recognition, and provides a method for unmanned aerial vehicle target trajectory recognition based on the RoPE-BiLSTM model, which can process different tasks through different networks under the premise of limited model size and data storage space, and realizes unmanned aerial vehicle target trajectory recognition with high generalization performance under the premise of significantly reducing the number of model parameters. The present invention also relates to a unmanned aerial vehicle target trajectory recognition system based on the RoPE-BiLSTM model.

[0007] The technical solution of the present invention is as follows:

[0008] A method for identifying UAV target trajectory based on RoPE-BiLSTM model, characterized by comprising the following steps:

[0009] Data acquisition and preprocessing steps: collecting trajectory points of multiple UAV targets and non-UAV targets within a certain time period at regular intervals, and then fitting multiple UAV trajectories and non-UAV trajectories, and then obtaining the trajectory data of each trajectory point in each trajectory to generate a trajectory data set, and dividing the trajectory data set into a training set and a test set according to a preset ratio and preprocessing; the trajectory data includes slant range, relative altitude, azimuth, radial velocity and radar scattering cross section;

[0010] Trajectory dynamic time series feature extraction steps: based on the pre-processed tilt moment, relative height and azimuth, the three-dimensional spatial coordinates of each trajectory point are calculated respectively; based on the three-dimensional spatial coordinates of each trajectory point, the relative displacement of each trajectory point in each trajectory in the horizontal coordinate direction, the vertical coordinate direction and the vertical coordinate direction respectively with the first trajectory point collected in the time period is calculated; then, based on the three-dimensional spatial coordinates of any two adjacent trajectory points and the time interval between the two trajectory points, the average speed, average acceleration, heading angle and pitch angle of the UAV target and non-UAV target between any two trajectory points are calculated respectively;

[0011] Trajectory static statistical feature extraction steps: calculate the average and standard deviation of the horizontal coordinates, vertical coordinates and vertical coordinates of all trajectory points in each trajectory respectively, and use the box counting method to calculate the box dimension of each trajectory based on the three-dimensional spatial coordinates of all trajectory points in each trajectory; then divide the space containing all trajectory points in each trajectory into several grid cells, and count the number of trajectory points falling in each grid cell, calculate the probability of occurrence of each grid cell in the entire trajectory based on the number of trajectory points in each grid cell and the total number of trajectory points, and calculate the Shannon entropy of each trajectory based on the occurrence probability and the total number of grid cells; calculate the approximate characteristic value of the radar cross section based on the radar cross section, heading angle and pitch angle;

[0012] Model construction steps: Use relative position encoding and bidirectional long short-term memory network to build a RoPE-BiLSTM model; then use the multi-layer perceptron model and adopt Dropout technology to build an improved multi-layer perceptron model; input the training set into the RoPE-BiLSTM model and the improved multi-layer perceptron model for training, and obtain the trained RoPE-BiLSTM model and the improved multi-layer perceptron model, and test and analyze them through the test set to obtain the final trained and tested RoPE-BiLSTM model and the improved multi-layer perceptron model;

[0013] Target trajectory identification steps: relative displacement, radar cross section, radial velocity, average velocity, average acceleration and heading angle are input as dynamic time series features into the trained and tested RoPE-BiLSTM model to obtain dynamic feature classification results; mean value, standard deviation, box dimension, Shannon entropy and approximate eigenvalue of radar cross section are input as static statistical features into the trained and tested improved multi-layer perceptron model to obtain static feature classification results; Bayesian linear layer is then used to identify the dynamic feature classification results and static feature classification results to identify UAV target trajectories and non-UAV target trajectories.

[0014] Preferably, in the target trajectory recognition step, the principal component analysis method is also used to reduce the dimension of the static statistical features, and the reduced-dimensional static statistical features are respectively input into the trained and tested improved multi-layer perceptron model to obtain the static feature classification results.

[0015] Preferably, in the data acquisition and preprocessing step, the preprocessing includes: firstly filtering the trajectory data using the 3-σ method, and then supplementing the missing values ​​in the filtered trajectory data using the interpolation method to ensure the continuity and integrity of the trajectory; and smoothing the supplemented trajectory data using the moving average method to obtain the smoothed trajectory data.

[0016] Preferably, in the data acquisition and preprocessing steps, random oversampling technology is also used to expand the number of drone targets in the training set to ensure that the model will not be affected by data imbalance during the training process.

[0017] Preferably, in the target trajectory recognition step, a normalization method is used to process the dynamic time series features and the static statistical features respectively, and then the normalized dynamic time series features are respectively input into the trained and tested RoPE-BiLSTM model to obtain the dynamic feature classification results; the normalized static statistical features are respectively input into the trained and tested improved multi-layer perceptron model to obtain the static feature classification results.

[0018] A UAV target trajectory recognition system based on the RoPE-BiLSTM model is characterized by comprising a data acquisition and preprocessing module, a trajectory dynamic time series feature extraction module, a trajectory static statistical feature extraction module, a model building module and a target trajectory recognition module connected in sequence.

[0019] The data acquisition and preprocessing module collects the trajectory points of multiple UAV targets and non-UAV targets within a certain time period at regular intervals, and then fits multiple UAV trajectories and non-UAV trajectories, and then obtains the trajectory data of each trajectory point in each trajectory to generate a trajectory data set, and divides the trajectory data set into a training set and a test set according to a preset ratio and performs preprocessing; the trajectory data includes slant range, relative altitude, azimuth, radial velocity and radar scattering cross section;

[0020] The trajectory dynamic time series feature extraction module calculates the three-dimensional spatial coordinates of each trajectory point based on the preprocessed tilt moment, relative height and azimuth, and calculates the relative displacement of each trajectory point in each trajectory in the horizontal coordinate direction, the vertical coordinate direction and the vertical coordinate direction with the first trajectory point collected in the time period according to the three-dimensional spatial coordinates of each trajectory point; and then calculates the average speed, average acceleration, heading angle and pitch angle of the UAV target and non-UAV target between any two trajectory points according to the three-dimensional spatial coordinates of any two adjacent trajectory points and the time interval between the two trajectory points;

[0021] The trajectory static statistical feature extraction module calculates the average and standard deviation of the horizontal coordinates, vertical coordinates and vertical coordinates of all trajectory points in each trajectory, and calculates the box dimension of each trajectory by using the box counting method based on the three-dimensional spatial coordinates of all trajectory points in each trajectory; then divides the space containing all trajectory points in each trajectory into a number of grid cells, and counts the number of trajectory points falling in each grid cell, calculates the occurrence probability of each grid cell in the entire trajectory according to the number of trajectory points in each grid cell and the total number of trajectory points, and calculates the Shannon entropy of each trajectory according to the occurrence probability and the total number of grid cells; calculates the approximate characteristic value of the radar cross section according to the radar cross section, heading angle and pitch angle;

[0022] The model building module uses relative position coding and bidirectional long short-term memory network to build a RoPE-BiLSTM model; then uses the multi-layer perceptron model and adopts Dropout technology to build an improved multi-layer perceptron model; the training set is input into the RoPE-BiLSTM model and the improved multi-layer perceptron model for training, respectively, to obtain the trained RoPE-BiLSTM model and the improved multi-layer perceptron model, and the test set is used for testing and analysis, to obtain the final trained and tested RoPE-BiLSTM model and the improved multi-layer perceptron model;

[0023] The target trajectory recognition module inputs relative displacement, radar cross section, radial velocity, average velocity, average acceleration and heading angle as dynamic time series features into the trained and tested RoPE-BiLSTM model to obtain dynamic feature classification results; and inputs average value, standard deviation, box dimension, Shannon entropy and approximate eigenvalue of radar cross section into the trained and tested improved multi-layer perceptron model as static statistical features to obtain static feature classification results; and then uses the Bayesian linear layer to identify the dynamic feature classification results and the static feature classification results to identify the UAV target trajectory and the non-UAV target trajectory.

[0024] Preferably, in the target trajectory recognition module, principal component analysis is also used to reduce the dimension of static statistical features, and the reduced-dimensional static statistical features are input into a trained and tested improved multi-layer perceptron model to obtain static feature classification results.

[0025] Preferably, in the data acquisition and preprocessing module, the preprocessing includes: firstly filtering the trajectory data using the 3-σ method, and then supplementing the missing values ​​in the filtered trajectory data using the interpolation method to ensure the continuity and integrity of the trajectory; and smoothing the supplemented trajectory data using the moving average method to obtain the smoothed trajectory data.

[0026] Preferably, in the data acquisition and preprocessing module, random oversampling technology is also used to expand the number of drone targets in the training set to ensure that the model will not be affected by data imbalance during the training process.

[0027] Preferably, in the target trajectory recognition module, a normalization method is used to process the dynamic time series features and the static statistical features respectively, and then the normalized dynamic time series features are input into the trained and tested RoPE-BiLSTM model to obtain the dynamic feature classification results; the normalized static statistical features are input into the trained and tested improved multi-layer perceptron model to obtain the static feature classification results.

[0028] The beneficial effects of the present invention are:

[0029] The present invention provides a method for identifying UAV target trajectories based on the RoPE-BiLSTM model. Based on the trajectory data of UAV targets and non-UAV targets, the position of the target trajectory is converted into a three-dimensional space coordinate, and a specific calculation method is used to calculate the dynamic time series characteristics of the trajectory and the static statistical characteristics of the trajectory. Then, relative position coding and a bidirectional long short-term memory network are used to construct an improved RoPE-BiLSTM model, a multi-layer perceptron model is used and the Dropout technology is used to construct an improved multi-layer perceptron model, and then the dynamic time series characteristics of the trajectory are input into the RoPE-BiLSTM model, and the static statistical characteristics of the trajectory are input into the improved multi-layer perceptron model. Different tasks are handled by different networks. Under the premise of significantly reducing the number of model parameters, a drone target trajectory recognition model with high generalization performance is realized. In addition, the relative displacement between the differential sampling trajectory points and the absolute position are separated through the RoPE-BiLSTM model, which avoids the loss of information in the normalization process of the data, thereby ensuring that the model accurately recognizes the drone trajectory, further strengthening the model's attention to long-term dependent features, making feature extraction more sufficient in the spatiotemporal dimension and adapting to diversified trajectory patterns; by improving the multi-layer perceptron model, deep static feature representation can be extracted, so that static features can be used for classification more efficiently, and the generalization performance of the model is improved. Finally, the Bayesian linear layer is used to identify the dynamic feature classification results and static feature classification results of the two models, respectively, to identify the drone target trajectory and non-drone target trajectory, which can achieve good generalization performance in the drone target trajectory recognition task of scene migration under the premise of limited model size and data storage space.

[0030] The present invention designs a model training paradigm for separating dynamic time series features from static statistical features for the task of unmanned aerial vehicle target trajectory recognition. The dynamic features between trajectory points within a trajectory are separated from the static features of the entire trajectory. For the spatial information of the trajectory points, only the dynamic features such as relative displacement, radial velocity, and average speed are input into the RoPE-BiLSTM model to extract dynamic behavior information. For the static statistical features such as the mean and variance of trajectory coordinates, box dimension, and Shannon entropy, feature extraction is performed through an improved multi-layer perceptron model. After that, the results output by the two models are feature fused through a Bayesian linear layer. The relative displacement between the differentially sampled trajectory points is separated from the absolute position to avoid information loss during the normalization process of the data, thereby ensuring that the model can accurately identify the trajectory of the unmanned aerial vehicle.

[0031] The present invention also relates to a UAV target trajectory recognition system based on the RoPE-BiLSTM model. The system corresponds to the above-mentioned UAV target trajectory recognition method based on the RoPE-BiLSTM model, and can be understood as a system for realizing the above-mentioned UAV target trajectory recognition method based on the RoPE-BiLSTM model, including a data acquisition and preprocessing module, a trajectory dynamic time series feature extraction module, a trajectory static statistical feature extraction module, a model building module and a target trajectory recognition module connected in sequence. The modules cooperate with each other and can process different tasks through different networks under the premise of limited model size and data storage space, thereby realizing UAV target trajectory recognition with high generalization performance while significantly reducing the amount of model parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flow chart of the UAV target trajectory recognition method based on the RoPE-BiLSTM model of the present invention.

[0033] Figure 2 It is a schematic diagram of the BiLSTM principle of the present invention.

[0034] Figure 3 It is a schematic diagram of the RoPE-attention principle of the present invention.

[0035] Figure 4 It is a flow chart of the target trajectory identification steps of the present invention. DETAILED DESCRIPTION

[0036] The present invention will be described below in conjunction with the accompanying drawings.

[0037] The present invention relates to a method for identifying drone target trajectories based on a RoPE-BiLSTM model, which can also be understood as a method for identifying drone target trajectories based on a RoPE-BiLSTM model and an improved multi-layer perceptron model. By constructing a RoPE-BiLSTM model for extracting time series data features and a multi-layer perceptron model for processing physical properties such as RCS, good generalization performance is achieved in the drone target trajectory identification task of scene migration under the premise of limited model size and data storage space. The flowchart of the method is shown in FIG. Figure 1 As shown, construct a radar feature data set: original data → radar data conversion → calculation of static and dynamic features → data cleaning → merging time series data; design network and train model; model comparison and optimization; drone target detection and classification recognition. The following is a specific description. The method includes the following steps in sequence:

[0038] Data acquisition and preprocessing steps: collect trajectory points of multiple UAV targets and non-UAV targets within a certain time period at regular intervals, and then fit multiple UAV trajectories and non-UAV trajectories, and then obtain the trajectory data of each trajectory point in each trajectory to generate a trajectory data set, and divide the trajectory data set into a training set and a test set according to a preset ratio and preprocess them; the trajectory data includes slant range, azimuth, relative altitude, radial velocity, relative time and radar scattering cross section.

[0039] Specifically, firstly, the trajectory points of multiple UAV targets and non-UAV targets in a certain time period are collected from the radar system at intervals (every 2s), and then all the trajectory points of each target are fitted to obtain multiple UAV trajectories and non-UAV trajectories, and then the trajectory data of each trajectory point in each trajectory is obtained to generate a trajectory data set, and all the trajectory data in the trajectory data set are divided into a training set and a test set according to a ratio of 9:1 and preprocessed; wherein, the length of each trajectory is determined according to the change of adjacent labels, and the specific trajectory length is uncertain; the total number of trajectories in the data set is 58,613, 41,639 non-UAV target trajectories and 16,974 UAV target trajectories. All UAV target trajectories and non-UAV target trajectories in the data set are labeled and classified into two labels, respectively: label "0" represents non-UAV targets; label "1" represents UAV targets. There are a total of 483,907 rows of records in the data set, and the length of trajectory data is different. Preferably, each trajectory data records six main features of the target: including slant range, azimuth, relative height between UAV target or non-UAV target and radar, radial velocity and radar scattering cross section.

[0040] Preferably, the 3-σ method is first used to filter the trajectory data. For the missing values ​​in the trajectory data, the interpolation method is used to supplement them to ensure the continuity and integrity of the trajectory; and in order to reduce the noise interference in the data, the moving average method is used to smooth the supplemented trajectory data to further improve the data quality. Among them, for some unrecoverable data, they are directly eliminated; for the abnormal values ​​in the characteristics such as flight altitude, slant range, speed, etc., they are eliminated by setting a reasonable threshold to remove those data that obviously do not conform to the movement law of drones or non-drone targets. In addition, since the number of drone data and flying drone data in the training set is seriously uneven, and the number of the two in the test set is similar, this will lead to an imbalance problem for drone and non-drone target data. Therefore, the random oversampling technology is used to expand the number of samples of drone targets in the training set to ensure that the model will not be affected by data imbalance during the training process. Before the subsequent data is input into the model, the minimum-maximum normalization method is used to map the data to [0, 1].

[0041] The steps of extracting dynamic time series features of trajectories are as follows: the three-dimensional spatial coordinates of each trajectory point are calculated based on the preprocessed tilt moment, relative height and azimuth, and the relative displacement of each trajectory point in each trajectory in the horizontal, vertical and vertical directions with the first trajectory point collected in the time period is calculated based on the three-dimensional spatial coordinates of each trajectory point; then, the dynamic features such as the average speed, average acceleration, heading angle and pitch angle of the UAV target and non-UAV target between any two trajectory points are calculated based on the three-dimensional spatial coordinates of any two adjacent trajectory points and the time interval between the two trajectory points. These features describe the target motion behavior and provide effective support for subsequent trajectory recognition.

[0042] Specifically, in order to facilitate the processing of subsequent models, the trajectory data needs to be converted into three-dimensional space coordinates (xyz coordinates), which provide a description of the position of the target in three-dimensional space. First, the three-dimensional space coordinates of each trajectory point are calculated based on the pre-processed slant moment d, relative height h, and azimuth θ, that is, the target position is converted into three-dimensional space coordinates to support the subsequent motion trajectory analysis, and the calculation is performed according to the following formula:

[0043]

[0044] In the above formula, d is the slant moment, h is the relative height between the UAV target or non-UAV target and the radar, and θ is the azimuth angle.

[0045] Then, according to the three-dimensional spatial coordinates of each trajectory point, the relative displacement of each trajectory point in each trajectory in the horizontal coordinate direction, the vertical coordinate direction and the vertical coordinate direction (i.e., in the x-axis direction, the y-axis direction and the z-axis direction) with the first trajectory point collected in the time period is calculated. That is, the coordinates of the trajectory point at the first sampling time point in a certain trajectory are set as (x 0 ,y 0 , z 0 ), then the relative displacement (x rel ,y rel , z rel ) is calculated according to the following formula:

[0046]

[0047] Then, according to the three-dimensional spatial coordinates of any two adjacent trajectory points and the time interval between the two trajectory points, the average velocity v, average acceleration a, heading angle φ and pitch angle θ of the UAV target and non-UAV target between any two trajectory points are calculated respectively according to the following formulas:

[0048]

[0049]

[0050] The steps of extracting static statistical features of trajectories are as follows: the mean and standard deviation of the horizontal, vertical and vertical coordinates of all trajectory points in each trajectory are calculated respectively, and the box dimension of each trajectory is calculated by the box counting method based on the three-dimensional spatial coordinates of all trajectory points in each trajectory; the space containing all trajectory points in each trajectory is divided into several grid cells, and the number of trajectory points falling in each grid cell is counted, and the probability of occurrence of each grid cell in the entire trajectory is calculated according to the number of trajectory points in each grid cell and the total number of trajectory points, and the Shannon entropy of each trajectory is calculated according to the occurrence probability and the total number of grid cells, and the box dimension of each trajectory and statistical indicators such as Shannon entropy are calculated according to the trajectory characteristics of the target space to measure the distribution characteristics of the trajectory points; the approximate characteristic value of the radar cross section is calculated according to the radar cross section, heading angle and pitch angle.

[0051] Specifically, first calculate the mean and standard deviation of the horizontal, vertical and horizontal coordinates of all trajectory points in each trajectory, that is, take the mean and variance of x, y and z of all trajectory points in each trajectory to extract the statistical properties of the target distribution in space. Then, based on the three-dimensional spatial coordinates of all trajectory points in each trajectory, the sampling box counting method is used to calculate the box dimension of each trajectory. Among them, the box dimension is a fractal dimension, which is estimated by the box counting method. The key idea is to measure the distribution pattern of the sequence by dividing the space and counting the number of boxes. The higher the value, the more complex the sequence. The one-dimensional time series is embedded into a high-dimensional phase space through delayed coordinates, and then the embedded phase space is divided into grids or "boxes", and the number of boxes containing at least the time series data points is counted. The calculation formula of the box dimension is:

[0052]

[0053] Among them, ε represents the side length of the grid or "box" (Box), and N(ε) represents the number of boxes that contain at least one time series data point.

[0054] Then divide the space containing all the trajectory points in each trajectory into several grid units (or discrete areas), and count the number of trajectory points falling in each grid unit. The probability of each grid unit appearing in the entire trajectory is calculated based on the number of trajectory points in each grid unit and the total number of trajectory points. The Shannon entropy of each trajectory is calculated based on the probability of appearance and the total number of grid units. Among them, the Shannon entropy measures the amount of information brought by the probability distribution of each state in the trajectory. The higher the entropy, the more uniform the state distribution in the trajectory and the greater the uncertainty of the trajectory; conversely, the lower the entropy, the more orderly and regular the trajectory. The formula is:

[0055]

[0056] Where X represents the state of the trajectory in the time series (i.e., the grid cell or discrete region in space), p(x i ) represents the probability of occurrence of the i-th state in the entire trajectory, and n is the total number of states.

[0057] Finally, the approximate characteristic value of the radar cross section RCS is calculated based on the radar cross section, heading angle and elevation angle. For the radar cross section RCS information, due to the limitation of the data set, the theoretical value cannot be accurately calculated by the field position. Therefore, the partial derivatives of RCS with respect to the azimuth and elevation angle are used to calculate the approximate characteristic value σ of the radar cross section RCS 0 , calculated according to the following formula:

[0058]

[0059] In the above formula, i represents the trajectory point.

[0060] Model construction steps: Use relative position encoding and bidirectional long short-term memory network to construct an improved RoPE-BiLSTM model; then use the multi-layer perceptron model and adopt Dropout technology to construct an improved multi-layer perceptron model; input the training set into the RoPE-BiLSTM model and the improved multi-layer perceptron model for training, and obtain the trained RoPE-BiLSTM model and the trained improved multi-layer perceptron model, and test and analyze the two models respectively through the test set, and obtain the final trained and tested RoPE-BiLSTM model and the trained and tested improved multi-layer perceptron model.

[0061] Specifically, the RoPE-BiLSTM model is first constructed using relative position encoding and bidirectional long short-term memory network. By designing and building a RoPE-BiLSTM model based on the bidirectional long short-term memory network (BiLSTM), the ability to capture bidirectional time-dependent information is effectively enhanced. At the same time, the RoPE-BiLSTM model introduces the relative position encoding (RoPE-attention) attention mechanism based on the traditional bidirectional long short-term memory network (BiLSTM) to enhance the model's sensitivity to dynamic changes in trajectories, effectively improving the model's ability to understand the position encoding of time series data. The BiLSTM network provides a more refined dynamic feature expression by capturing the bidirectional correlation between the front and back of the drone's trajectory, while the RoPE-attention mechanism further strengthens the model's attention to long-term dependent features, making feature extraction more sufficient in the spatiotemporal dimension and adapting to diverse trajectory patterns.

[0062] In the actual modeling process, by adding multiple layers of LSTM networks in BiLSTM, the output of the previous LSTM network is used as the input of the next LSTM network, and multiple layers of LSTM networks are stacked to ensure that deep abstract features can be extracted and the nonlinear fitting ability of the model can be increased. BiLSTM processes information flow from two directions (forward and reverse), and it considers both the forward and reverse information of the input sequence at the same time, which can capture sequence features more comprehensively. Compared with traditional machine learning algorithms, BiLSTM has stronger sequence information processing capabilities. Its structure is as follows: Figure 2 As shown in the figure, the calculation process of the LSTM network can be summarized as follows: by forgetting the information in the cell state and memorizing new information, the information useful for subsequent moment calculations can be transmitted, while the useless information is discarded, and the hidden state h is output at each time step. t Among them, forgetting, memory and output are determined by the hidden state h of the previous moment. t-1 and the current input x t The calculated forget gate f t , Memory Gate t , output gate o t to control.

[0063] In order to better capture the relative position information between word vectors, the RoPE method is used to achieve relative position encoding through the rotation matrix. The core idea of ​​RoPE is to apply a rotation transformation to the word embedding vector to inject position information. For each position in the sequence, RoPE adjusts the word vector through a rotation matrix, and the angle of rotation is related to the position index. This rotation operation not only retains the modulus of the vector (that is, the size of the word vector remains unchanged), but also introduces relative position information. RoPE proposes a new position encoding function based on the rotation matrix, such as Figure 3 The improved attention mechanism of RoPE-attention shown in Figure 1. By introducing rotation position encoding, some limitations of traditional position encoding methods are solved, especially the effectiveness of long sequence modeling, and the model's understanding of position encoding of time series data is increased. For the word vector v at position p, its position encoding is represented as the result of a rotation matrix acting on the word vector. Let p be the absolute position of the word and θ be a rotation angle related to the position. Then the position encoding function of RoPE can be expressed as:

[0064] v p =R(θ p ) (7)

[0065] In the above formula, θ p is the rotation angle corresponding to position p, which is usually linearly related to position p. This means that RoPE adds absolute position information to word vectors by rotating them, while implicitly capturing relative position information.

[0066] Among them, x 1 , x 2 The input data point can be a word embedding vector in a text sequence or a pixel value in an image; by applying the matrix m θ , the original input data point x 1 , x 2 Transform to the new position x' 1 , x' 2 (i.e. the transformed output data point); Query and Key are key components of the attention mechanism, which are used to calculate the correlation between different data points. Position Embedding is a method of adding position information to the query and key, so that the model can understand the relative position relationship between data points. Position is the position information of the input data point.

[0067] Then, the multi-layer perceptron model is used and the Dropout technology is used to construct an improved multi-layer perceptron model. Among them, the drone trajectory often contains a large number of low-frequency static features that do not change over time, but are of great significance to the description of the trajectory pattern. In order to refine the static feature information, the principal component analysis (PCA) is first used to reduce the dimension of the high-dimensional static features in the trajectory to reduce redundancy and remove noise, thereby ensuring that the most discriminative information is retained. The reduced-dimensional static features are further input into the improved multi-layer perceptron (MLP) with dropout regularization to extract deep-level static feature representations, so that the static features can be used more efficiently for classification and improve the generalization performance of the model. The overall structure of the constructed RoPE-BiLSTM model and the improved multi-layer perceptron model is as follows: Figure 4 As shown. Finally, the training set is input into the RoPE-BiLSTM model and the improved multi-layer perceptron model for training, and the trained RoPE-BiLSTM model and the improved multi-layer perceptron model are obtained. The trained RoPE-BiLSTM model and the improved multi-layer perceptron model are tested and analyzed through the test set, and the final trained and tested RoPE-BiLSTM model and the improved multi-layer perceptron model are obtained.

[0068] Target trajectory recognition steps: Figure 4As shown in the figure, the relative displacement in x, y, and z dimensions, radar cross section RCS, radial rate, relative time (i.e., the relative time between each trajectory point and the first trajectory point in each trajectory), average speed, average acceleration, and heading angle are used as trajectory dynamic time series features (referred to as dynamic time series features), and the dynamic time series features are processed by normalization method. The normalized dynamic time series features are input into the trained and tested RoPE-BiLSTM model respectively. The RoPE-attention mechanism of the RoPE-BiLSTM model is used to increase the model's understanding of the time series data position encoder, and the BiLSTM layer is used to capture the bidirectional correlation between the front and back of the UAV trajectory, providing a more refined dynamic feature expression, and obtaining the dynamic feature classification result. The x, y, The z-coordinate mean, standard deviation, box dimension, Shannon entropy and approximate eigenvalues ​​of the radar cross section are used as static statistical features of the trajectory (referred to as static statistical features). The principal component analysis method is used to reduce the dimension of the static statistical features, and the normalization method is used to process the static statistical features after dimensionality reduction. The normalized static statistical features are then input into the trained and tested improved multi-layer perceptron model respectively. This model can be understood as a multi-layer perceptron + dropout technology, that is, an improved multi-layer perceptron (MLP) with dropout regularization that has been trained and tested. The deep-level static feature representation is extracted to obtain the static feature classification result. The Bayesian linear layer is then used to identify the dynamic feature classification results and the static feature classification results to identify the UAV target trajectory and the non-UAV target trajectory.

[0069] The present invention also relates to a UAV target trajectory recognition system based on the RoPE-BiLSTM model. The system corresponds to the UAV target trajectory recognition method based on the RoPE-BiLSTM model, and can be understood as a system for implementing the above method. The system includes a data acquisition and preprocessing module, a trajectory dynamic time series feature extraction module, a trajectory static statistical feature extraction module, a model construction module and a target trajectory recognition module connected in sequence. Specifically,

[0070] The data acquisition and preprocessing module collects the trajectory points of multiple UAV targets and non-UAV targets within a certain time period at regular intervals, and then fits multiple UAV trajectories and non-UAV trajectories, and then obtains the trajectory data of each trajectory point in each trajectory to generate a trajectory data set, and divides the trajectory data set into a training set and a test set according to a preset ratio and performs preprocessing; the trajectory data includes slant range, relative altitude, azimuth, radial velocity and radar scattering cross section;

[0071] The trajectory dynamic time series feature extraction module calculates the three-dimensional spatial coordinates of each trajectory point based on the preprocessed tilt moment, relative height and azimuth, and calculates the relative displacement of each trajectory point in each trajectory in the horizontal coordinate direction, the vertical coordinate direction and the vertical coordinate direction with the first trajectory point collected in the time period according to the three-dimensional spatial coordinates of each trajectory point; and then calculates the average speed, average acceleration, heading angle and pitch angle of the UAV target and non-UAV target between any two trajectory points according to the three-dimensional spatial coordinates of any two adjacent trajectory points and the time interval between the two trajectory points;

[0072] The trajectory static statistical feature extraction module calculates the average and standard deviation of the horizontal coordinates, vertical coordinates and vertical coordinates of all trajectory points in each trajectory, and calculates the box dimension of each trajectory by using the box counting method based on the three-dimensional spatial coordinates of all trajectory points in each trajectory; then divides the space containing all trajectory points in each trajectory into a number of grid cells, and counts the number of trajectory points falling in each grid cell, calculates the occurrence probability of each grid cell in the entire trajectory according to the number of trajectory points in each grid cell and the total number of trajectory points, and calculates the Shannon entropy of each trajectory according to the occurrence probability and the total number of grid cells; calculates the approximate characteristic value of the radar cross section according to the radar cross section, heading angle and pitch angle;

[0073] The model building module uses relative position coding and bidirectional long short-term memory network to build a RoPE-BiLSTM model; then uses the multi-layer perceptron model and adopts Dropout technology to build an improved multi-layer perceptron model; the training set is input into the RoPE-BiLSTM model and the improved multi-layer perceptron model for training, respectively, to obtain the trained RoPE-BiLSTM model and the improved multi-layer perceptron model, and the test set is used for testing and analysis, to obtain the final trained and tested RoPE-BiLSTM model and the improved multi-layer perceptron model;

[0074] The target trajectory recognition module inputs relative displacement, radar cross section, radial velocity, average velocity, average acceleration and heading angle as dynamic time series features into the trained and tested RoPE-BiLSTM model to obtain dynamic feature classification results; and inputs average value, standard deviation, box dimension, Shannon entropy and approximate eigenvalue of radar cross section into the trained and tested improved multi-layer perceptron model as static statistical features to obtain static feature classification results; and then uses the Bayesian linear layer to identify the dynamic feature classification results and the static feature classification results to identify the UAV target trajectory and the non-UAV target trajectory.

[0075] Preferably, in the target trajectory recognition module, the principal component analysis method is also used to reduce the dimension of the static statistical features, and the reduced-dimensional static statistical features are input into the trained and tested improved multi-layer perceptron model to obtain the static feature classification results.

[0076] Preferably, in the data acquisition and preprocessing module, the preprocessing includes: firstly using the 3-σ method to filter the trajectory data, and then using the interpolation method to supplement the missing values ​​in the filtered trajectory data to ensure the continuity and integrity of the trajectory; and using the moving average method to smooth the supplemented trajectory data to obtain the smoothed trajectory data.

[0077] Preferably, in the data acquisition and preprocessing module, random oversampling technology is also used to expand the number of drone targets in the training set to ensure that the model will not be affected by data imbalance during the training process.

[0078] Preferably, in the target trajectory recognition module, a normalization method is used to process the dynamic time series features and the static statistical features respectively, and then the normalized dynamic time series features are input into the trained and tested RoPE-BiLSTM model to obtain the dynamic feature classification results; the normalized static statistical features are input into the trained and tested improved multi-layer perceptron model to obtain the static feature classification results.

[0079] The present invention provides an objective and scientific UAV target trajectory recognition method and system based on the RoPE-BiLSTM model. The method separates the dynamic features between trajectory points in a trajectory from the static features of the entire trajectory. For the spatial information of the trajectory points, only the dynamic features are input into the RoPE-BiLSTM model to extract dynamic behavior information. For the static statistical features, feature extraction is performed through an improved multi-layer perceptron model, and then the results output by the two models are feature fused through a Bayesian linear layer. The relative displacement between the differential sampling trajectory points is separated from the absolute position to avoid information loss in the normalization process of the data, thereby ensuring that the model accurately recognizes the UAV trajectory.

[0080] It should be noted that the above-described specific implementations can enable those skilled in the art to more fully understand the invention, but do not limit the invention in any way. Therefore, although this specification has described the invention in detail with reference to the drawings and embodiments, those skilled in the art should understand that the invention can still be modified or replaced by equivalents. In short, all technical solutions and improvements that do not deviate from the spirit and scope of the invention should be included in the protection scope of the patent for the invention.

Claims

1. A method for identifying UAV target trajectories based on the RoPE-BiLSTM model, characterized in that: The following steps are involved: Data acquisition and preprocessing steps: collecting trajectory points of multiple UAV targets and non-UAV targets within a certain time period at regular intervals, and then fitting multiple UAV trajectories and non-UAV trajectories, and then obtaining the trajectory data of each trajectory point in each trajectory to generate a trajectory data set, and dividing the trajectory data set into a training set and a test set according to a preset ratio and preprocessing; the trajectory data includes slant range, relative altitude, azimuth, radial velocity and radar scattering cross section; Trajectory dynamic time series feature extraction steps: based on the pre-processed tilt moment, relative height and azimuth, the three-dimensional spatial coordinates of each trajectory point are calculated respectively; based on the three-dimensional spatial coordinates of each trajectory point, the relative displacement of each trajectory point in each trajectory in the horizontal coordinate direction, the vertical coordinate direction and the vertical coordinate direction respectively with the first trajectory point collected in the time period is calculated; then, based on the three-dimensional spatial coordinates of any two adjacent trajectory points and the time interval between the two trajectory points, the average speed, average acceleration, heading angle and pitch angle of the UAV target and non-UAV target between any two trajectory points are calculated respectively; Trajectory static statistical feature extraction steps: calculate the average and standard deviation of the horizontal coordinates, vertical coordinates and vertical coordinates of all trajectory points in each trajectory respectively, and use the box counting method to calculate the box dimension of each trajectory based on the three-dimensional spatial coordinates of all trajectory points in each trajectory; then divide the space containing all trajectory points in each trajectory into several grid cells, and count the number of trajectory points falling in each grid cell, calculate the probability of occurrence of each grid cell in the entire trajectory based on the number of trajectory points in each grid cell and the total number of trajectory points, and calculate the Shannon entropy of each trajectory based on the occurrence probability and the total number of grid cells; calculate the approximate characteristic value of the radar cross section based on the radar cross section, heading angle and pitch angle; Model construction steps: Use relative position encoding and bidirectional long short-term memory network to build a RoPE-BiLSTM model; then use the multi-layer perceptron model and adopt Dropout technology to build an improved multi-layer perceptron model; input the training set into the RoPE-BiLSTM model and the improved multi-layer perceptron model for training, and obtain the trained RoPE-BiLSTM model and the improved multi-layer perceptron model, and test and analyze them through the test set to obtain the final trained and tested RoPE-BiLSTM model and the improved multi-layer perceptron model; Target trajectory identification steps: relative displacement, radar cross section, radial velocity, average velocity, average acceleration and heading angle are input as dynamic time series features into the trained and tested RoPE-BiLSTM model to obtain dynamic feature classification results; mean value, standard deviation, box dimension, Shannon entropy and approximate eigenvalue of radar cross section are input as static statistical features into the trained and tested improved multi-layer perceptron model to obtain static feature classification results; Bayesian linear layer is then used to identify the dynamic feature classification results and static feature classification results to identify UAV target trajectories and non-UAV target trajectories.

2. The method for identifying unmanned aerial vehicle target trajectories based on the RoPE-BiLSTM model according to claim 1, characterized in that: In the target trajectory recognition step, the principal component analysis method is also used to reduce the dimension of the static statistical features, and the reduced static statistical features are respectively input into the trained and tested improved multi-layer perceptron model to obtain the static feature classification results.

3. The method for identifying unmanned aerial vehicle target trajectories based on the RoPE-BiLSTM model according to claim 1, characterized in that: In the data acquisition and preprocessing steps, the preprocessing includes: firstly using the 3-σ method to filter the trajectory data, then using the interpolation method to supplement the missing values ​​in the filtered trajectory data to ensure the continuity and integrity of the trajectory; and using the moving average method to smooth the supplemented trajectory data to obtain the smoothed trajectory data.

4. The method for identifying unmanned aerial vehicle target trajectories based on the RoPE-BiLSTM model according to claim 1, characterized in that: In the data acquisition and preprocessing steps, random oversampling technology is also used to expand the number of drone targets in the training set to ensure that the model will not be affected by data imbalance during the training process.

5. The method for identifying unmanned aerial vehicle target trajectories based on the RoPE-BiLSTM model according to claim 2 is characterized in that: In the target trajectory recognition step, a normalization method is used to process the dynamic time series features and the static statistical features respectively, and then the normalized dynamic time series features are respectively input into the trained and tested RoPE-BiLSTM model to obtain the dynamic feature classification results; the normalized static statistical features are respectively input into the trained and tested improved multi-layer perceptron model to obtain the static feature classification results.

6. A UAV target trajectory recognition system based on RoPE-BiLSTM model, characterized in that: It includes a data acquisition and preprocessing module, a trajectory dynamic time series feature extraction module, a trajectory static statistical feature extraction module, a model building module and a target trajectory recognition module, which are connected in sequence. The data acquisition and preprocessing module collects the trajectory points of multiple UAV targets and non-UAV targets within a certain time period at regular intervals, and then fits multiple UAV trajectories and non-UAV trajectories, and then obtains the trajectory data of each trajectory point in each trajectory to generate a trajectory data set, and divides the trajectory data set into a training set and a test set according to a preset ratio and performs preprocessing; the trajectory data includes slant range, relative altitude, azimuth, radial velocity and radar scattering cross section; The trajectory dynamic time series feature extraction module calculates the three-dimensional spatial coordinates of each trajectory point based on the preprocessed tilt moment, relative height and azimuth, and calculates the relative displacement of each trajectory point in each trajectory in the horizontal coordinate direction, the vertical coordinate direction and the vertical coordinate direction with the first trajectory point collected in the time period according to the three-dimensional spatial coordinates of each trajectory point; and then calculates the average speed, average acceleration, heading angle and pitch angle of the UAV target and non-UAV target between any two trajectory points according to the three-dimensional spatial coordinates of any two adjacent trajectory points and the time interval between the two trajectory points; The trajectory static statistical feature extraction module calculates the average value and standard deviation of the horizontal coordinate, the vertical coordinate and the ordinate of all trajectory points in each trajectory, and calculates the box dimension of each trajectory by using the box counting method based on the three-dimensional spatial coordinates of all trajectory points in each trajectory; Then divide the space containing all the track points in each track into several grid cells, and count the number of track points falling in each grid cell. According to the number of track points in each grid cell and the total number of track points, calculate the probability of each grid cell appearing in the entire track. According to the probability of appearance and the total number of grid cells, calculate the Shannon entropy of each track. Calculate the approximate characteristic value of the radar cross section according to the radar cross section, heading angle and pitch angle. The model building module uses relative position coding and bidirectional long short-term memory network to build a RoPE-BiLSTM model; then uses the multi-layer perceptron model and adopts Dropout technology to build an improved multi-layer perceptron model; the training set is input into the RoPE-BiLSTM model and the improved multi-layer perceptron model for training, respectively, to obtain the trained RoPE-BiLSTM model and the improved multi-layer perceptron model, and the test set is used for testing and analysis, to obtain the final trained and tested RoPE-BiLSTM model and the improved multi-layer perceptron model; The target trajectory recognition module inputs relative displacement, radar cross section, radial velocity, average velocity, average acceleration and heading angle as dynamic time series features into the trained and tested RoPE-BiLSTM model to obtain dynamic feature classification results; The mean value, standard deviation, box dimension, Shannon entropy and approximate characteristic value of radar cross section are input into the trained and tested improved multi-layer perceptron model as static statistical features to obtain the static feature classification results. The Bayesian linear layer is then used to identify the dynamic feature classification results and the static feature classification results to identify the UAV target trajectory and non-UAV target trajectory.

7. The UAV target trajectory recognition system based on the RoPE-BiLSTM model according to claim 6 is characterized in that: In the target trajectory recognition module, the principal component analysis method is also used to reduce the dimension of the static statistical features, and the reduced-dimensional static statistical features are input into the trained and tested improved multi-layer perceptron model to obtain the static feature classification results.

8. The UAV target trajectory recognition system based on the RoPE-BiLSTM model according to claim 6 is characterized in that: In the data acquisition and preprocessing module, the preprocessing includes: firstly using the 3-σ method to filter the trajectory data, then using the interpolation method to supplement the missing values ​​in the filtered trajectory data to ensure the continuity and integrity of the trajectory; and using the moving average method to smooth the supplemented trajectory data to obtain the smoothed trajectory data.

9. The UAV target trajectory recognition system based on the RoPE-BiLSTM model according to claim 6 is characterized in that: In the data acquisition and preprocessing module, random oversampling technology is also used to expand the number of drone targets in the training set to ensure that the model will not be affected by data imbalance during the training process.

10. The UAV target trajectory recognition system based on the RoPE-BiLSTM model according to claim 7, characterized in that: In the target trajectory recognition module, a normalization method is used to process the dynamic time series features and the static statistical features respectively, and then the normalized dynamic time series features are input into the trained and tested RoPE-BiLSTM model to obtain the dynamic feature classification results; The normalized static statistical features are input into the trained and tested improved multi-layer perceptron model to obtain the static feature classification results.

Images