An intention recognition method based on three-dimensional trajectory analysis of a drone cluster
By preprocessing and deep learning the 3D trajectory data of UAV swarms, and combining the improved HDBSCAN algorithm and graph neural network, the problems of accuracy and real-time performance in UAV swarm intent recognition in complex environments are solved, and efficient and accurate recognition of UAV swarm intent is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENYANG AEROSPACE UNIVERSITY
- Filing Date
- 2024-11-18
- Publication Date
- 2026-06-23
AI Technical Summary
Existing methods for identifying the intent of drone swarms struggle to achieve rapid and accurate identification in complex and dynamically changing combat environments. In particular, in multi-target cooperative combat scenarios, existing methods suffer from insufficient adaptability, high computational overhead, and poor real-time performance of multi-source information fusion.
By preprocessing, clustering analysis, and deep learning of the 3D trajectory data of drone swarms, and utilizing the improved HDBSCAN algorithm and graph neural network combined with a bidirectional GRU network, the 3D trajectory features of drone swarms are extracted and analyzed, enabling efficient identification of drone swarm intentions.
It improves the accuracy and reliability of drone swarm intent recognition, can dynamically adjust the clustering structure in complex scenarios, capture the time dependency of trajectories, and improve the accuracy and performance of recognition results.
Smart Images
Figure CN119557673B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of group behavior recognition of unmanned aerial vehicle (UAV) swarm, and in particular to an intention recognition method based on three-dimensional trajectory analysis of UAV swarm. BACKGROUND
[0002] With the rapid development of UAV technology, especially in the military field, the tactical application of UAV swarm is increasing. UAVs can work together to perform complex tasks and provide higher operational flexibility. However, as the complexity of the operational environment increases, it is difficult for commanders to analyze the behavior of UAV swarm in real time and recognize its intention, which cannot meet the requirements of timely and accurate decision-making, and also affects the rapid and accurate response of existing weapon systems. Therefore, there is an urgent need for an intention recognition method suitable for UAV swarm trajectory analysis.
[0003] In recent years, with the rapid expansion of UAV applications, UAV intention recognition methods have gradually become a research hotspot and received widespread attention. The purpose of UAV intention recognition is to infer its current or future tasks and actions by analyzing its behavior, trajectory and other related information, thereby providing decision support for defense and control. In this process, accurately recognizing the intention of UAV is crucial, as it directly affects the response speed and accuracy of the defense system. Existing intention recognition methods include rule-based methods, statistical learning-based methods and deep learning-based methods. However, these methods face many challenges, mainly in the following aspects: (1) Rule-based methods rely on manually set parameters and rules, lack self-adaptive ability, and are difficult to cope with complex and dynamic environments; (2) Statistical learning-based methods, such as SVM and random forest, can learn from data, but feature engineering is complex and difficult to handle multi-dimensional high-complexity data; (3) These methods mostly need to integrate multi-source information, such as radar data, electronic jamming state, communication information, etc., to achieve more comprehensive intention recognition, but in the process of multi-source information fusion, data processing and synchronization often bring a large computational overhead, affecting real-time performance and system robustness; (4) With the increasing demand for group cooperative combat of UAVs, group intention recognition becomes more complex, and the problem of multi-target cooperation has not been completely solved. How to extract accurate group intention from the trajectories of multi-UAV interaction is a difficult problem that needs to be solved urgently. SUMMARY
[0004] In view of the deficiencies of the prior art, the present application provides an intention recognition method based on three-dimensional trajectory analysis of UAV swarm, which uses data preprocessing, clustering analysis, deep learning and other methods to solve the problem of quickly and efficiently recognizing the intention of UAV swarm in war.
[0005] An intention recognition method based on three-dimensional trajectory analysis of UAV swarm, comprising the following steps:
[0006] S1: Obtain the 3D trajectory dataset of the UAV cluster and preprocess the 3D trajectories to obtain the preprocessed 3D trajectories; the 3D trajectory dataset of the UAV cluster includes the 3D trajectories of multiple UAVs, and each UAV's 3D trajectory includes several time steps and the 3D spatial coordinates of the trajectory points at that time step;
[0007] S1.1: Obtain the 3D trajectories of drones in the drone cluster, obtain the 3D trajectory dataset of the drone cluster, and filter the obtained 3D trajectories, deleting incomplete and abnormal 3D trajectories.
[0008] drones The three-dimensional trajectory is:
[0009] (1)
[0010] in, Indicates drone The three-dimensional trajectory This represents the time step, and 'i' represents the number of the trajectory point. Indicates the drone at time step The three-dimensional spatial coordinates of the trajectory points below, Indicates the number of trajectory points;
[0011] Set the minimum threshold for the length of the three-dimensional trajectory as follows: If the length of the three-dimensional trajectory If the three-dimensional trajectory is incomplete, it is considered incomplete and deleted; calculate the average velocity of the three-dimensional trajectory. If the average velocity of a certain three-dimensional trajectory exceeds a reasonable range If the three-dimensional trajectory is incorrect or abnormal, then it is considered to contain errors or anomalies. If so, then delete the 3D trajectory;
[0012] S1.2: Use cubic spline interpolation to smooth the three-dimensional trajectory and unify the three-dimensional trajectory to the same time dimension;
[0013] Specifically, this involves constructing a cubic spline function based on the selected 3D trajectory. ,in The first step is used to interpolate missing trajectory points; then, a uniform resampling time step is applied. Resampling is performed to generate trajectory points at the same time step;
[0014] The resampled trajectory points are represented as follows:
[0015] (3)
[0016] in, denotes the resampled trajectory point, and are the start time and end time of the three-dimensional trajectory, respectively;
[0017] S1.3: Standardize all three-dimensional trajectories to have the same dimension and range;
[0018] S1.4: Map the three-dimensional trajectories to the same geodetic map through coordinate transformation;
[0019] For the three-dimensional space coordinates of each trajectory point in the three-dimensional trajectory , it is mapped to the coordinate system of the geodetic map through the following matrix transformation formula:
[0020] (4)
[0021] wherein, is the transformation matrix, is the offset vector, is the three-dimensional space coordinates of the trajectory point mapped to the geodetic map;
[0022] S2: Cluster analysis is performed on the preprocessed three-dimensional trajectories to divide them into different unmanned aerial vehicle trajectory clusters;
[0023] S2.1: Extract trajectory points with large direction changes in the three-dimensional trajectory by calculating the changes in the angle of the three-dimensional trajectory;
[0024] Suppose that a three-dimensional trajectory is composed of trajectory points, and the three-dimensional space coordinates of the trajectory points in the three-dimensional trajectory are denoted as ;
[0025] First, for each two adjacent trajectory points on the three-dimensional trajectory, a vector is constructed from one trajectory point to the direction of the adjacent trajectory point:
[0026] (5)
[0027] wherein, is the three-dimensional space coordinates of the adjacent trajectory point;
[0028] For the two adjacent vectors and , the included angle between them is calculated through the dot product formula of the vector:
[0029] (6)
[0030] wherein, denotes the dot product of the two vectors, and It is the magnitude of the vector, and the formula is:
[0031] (7)
[0032] For every two adjacent trajectory points on the three-dimensional trajectory, calculate the difference in their motion direction. The specific calculation formula is as follows:
[0033] (8)
[0034] in, This indicates the difference in the direction of motion between two adjacent trajectory points. The value range is from 0 to 2. A larger value indicates a greater change in the direction of the 3D trajectory. The threshold should be set according to specific requirements. ,when At that time, the trajectory point is recorded, thereby extracting the trajectory points with large changes in direction in the three-dimensional trajectory;
[0035] S2.2: Calculate the similarity measure between three-dimensional trajectories using the starting trajectory point, ending trajectory point, and extracted trajectory points with large direction changes in the three-dimensional trajectory;
[0036] The two three-dimensional trajectories are expressed as follows:
[0037] (9)
[0038] (10)
[0039] in, and There are two three-dimensional trajectories. and Representing three-dimensional trajectories respectively and three-dimensional trajectory Trajectory points in; j are the numbers of the trajectory points. It is a three-dimensional trajectory The total number of starting trajectory points, ending trajectory points, and extracted trajectory points with significant changes in trajectory direction; It is a three-dimensional trajectory The total number of starting trajectory points, ending trajectory points, and extracted trajectory points with significant changes in trajectory direction;
[0040] First, calculate the three-dimensional trajectory separately. From the starting point, ending point, and points with significant direction changes on the trajectory to the three-dimensional trajectory Calculate the distance to the nearest trajectory point, and then calculate the 3D trajectory. From the starting point, ending point, and points with significant direction changes on the trajectory to the three-dimensional trajectory The average distance to the nearest trajectory point is used to obtain the 3D trajectory using the same method. From the starting point, ending point, and points with significant direction changes on the trajectory to the three-dimensional trajectory The mean distance to the nearest trajectory point is taken, and then the minimum of these two values is taken as the 3D trajectory. and three-dimensional trajectory The similarity of the three-dimensional trajectory shapes between them is calculated using the following formula:
[0041] (11)
[0042] in, Representing a three-dimensional trajectory and three-dimensional trajectory The similarity of the three-dimensional trajectory shapes between them; Representing a three-dimensional trajectory trajectory points on To three-dimensional trajectory Distance to the nearest trajectory point Representing a three-dimensional trajectory trajectory points on To three-dimensional trajectory Distance to the nearest trajectory point;
[0043] Secondly, for three-dimensional trajectories Calculate the relationship between the starting trajectory point, ending trajectory point, and trajectory points with large direction changes on the trajectory and the three-dimensional trajectory. The mean of the differences in the motion directions of the nearest trajectory points is used to similarly calculate the three-dimensional trajectory. The starting point, ending point, and points with significant changes in direction of the trajectory are compared with the three-dimensional trajectory. The mean of the differences in the motion directions of the nearest trajectory points is taken as the minimum of these two values, and the formula for calculating the three-dimensional trajectory direction similarity is as follows:
[0044] (12)
[0045] in, Representing a three-dimensional trajectory and three-dimensional trajectory The similarity of the three-dimensional trajectory directions between them; Representing a three-dimensional trajectory Upper trajectory point and three-dimensional trajectory Upper trajectory point The angle between the directions of motion, take The value is used to characterize the difference in the direction of motion;
[0046] Finally, the similarity between the three-dimensional trajectory shape and the three-dimensional trajectory direction is normalized and then combined to calculate the similarity measure between the three-dimensional trajectories. The calculation formula is as follows:
[0047] (13)
[0048] in, This represents the maximum value of the similarity between the three-dimensional trajectory shapes. This is a similarity measure between three-dimensional trajectories, with values ranging from... ;
[0049] S2.3: Based on the similarity measure between the three-dimensional trajectories, the improved HDBSCAN algorithm is used to perform cluster analysis on the preprocessed three-dimensional trajectories to obtain several UAV trajectory clusters;
[0050] The improved HDBSCAN algorithm replaces the core distance in the original HDBSCAN algorithm with the trajectory core distance. The original HDBSCAN algorithm's mutual reachability metric distance is improved to a trajectory mutual reachability metric distance. ;
[0051] The distance of the trajectory core The calculation method is as follows: calculate the custom distance function value between every two 3D trajectories. For a 3D trajectory... The custom distance function values between the trajectories and other 3D trajectories are arranged in ascending order, and the k-th custom distance function value is taken as the core distance of the trajectory. ;
[0052] The formula for calculating the custom distance function is as follows:
[0053] (14)
[0054] in, For three-dimensional trajectory and three-dimensional trajectory Custom distance function values between them;
[0055] The distance between the mutually reachable trajectories is calculated using the following formula:
[0056] (15)
[0057] in, This represents the distance between two 3D trajectories that are mutually reachable.
[0058] S3: Input each drone trajectory cluster into a pre-trained deep learning intent recognition network to perform intent recognition and obtain the intent prediction result for each drone trajectory cluster;
[0059] S3.1: Clustering a specific UAV trajectory cluster All the 3D trajectories are input into the input layer, where, , The number of 3D trajectories in the UAV trajectory cluster. For time step;
[0060] S3.2: Construct a graph structure based on the 3D trajectories in the UAV trajectory cluster;
[0061] S3.2.1: Treat each UAV corresponding to a 3D trajectory in the UAV trajectory cluster as a node, and initialize the node features for each node. Initialize as a feature vector And construct the initial feature matrix. ;
[0062] The feature vector Represents a node The initial features, where It is a node The mean of all three-dimensional spatial coordinates in the corresponding three-dimensional trajectory of the UAV;
[0063] S3.2.2: Construct a weighted adjacency matrix based on the similarity measure between the three-dimensional trajectories of the UAVs. ;
[0064] Specifically, it involves: a similarity measurement based on three-dimensional trajectories. Construct a similarity matrix , the elements Set similarity threshold ,like Not less than the threshold Then the node and nodes There are edges between them, weighted adjacency matrix elements in This is a similarity measure between 3D trajectories; if there are no edges, then the element... =0;
[0065] (16)
[0066] S3.2.3: In the weighted adjacency matrix Adding self-loops to the matrix yields a weighted adjacency matrix with added self-loops.
[0067] The adjacency matrix after adding the self-loop is: ,in It is the identity matrix;
[0068] S3.2.4: Based on the initial characteristic matrix and the adjacency matrix after adding self-loops, the graph structure is obtained;
[0069] S3.3: Input the graph structure into a two-layer graph neural network, and realize feature sharing within the UAV trajectory cluster through the two-layer graph neural network. That is, the feature vector of each node is aggregated with the feature vectors of its neighboring nodes to generate group structure features.
[0070] In each layer of the graph convolutional network, the node feature vector update follows the formula:
[0071] (17)
[0072] in, It is the first The feature matrix of the layer, It is the first The feature matrix of the layer, where each row represents the feature vector of a node. For the initial layer, This is the initial feature matrix; yes The degree matrix, whose diagonal elements , These are the elements in the weighted adjacency matrix after adding self-loops; It is the first Layer weight matrix; It is a non-linear activation function, ReLU.
[0073] S3.4: Input the three-dimensional trajectory in the UAV trajectory cluster into the bidirectional GRU network, extract the time series feature sequence set of each three-dimensional trajectory in the UAV trajectory cluster, and then obtain the time feature vector of the UAV trajectory cluster;
[0074] (20)
[0075] in, Indicates time step Drone The three-dimensional trajectory is in the output of the bidirectional GRU. In time step The input 3D trajectory, It is the output of the forward GRU. It is the output of the backward GRU;
[0076] Finally, based on the output of the bidirectional GRU (Bi-GRU) at each time step in the three-dimensional trajectory, the UAV is obtained. Time series feature set , The number of time steps is used to obtain the temporal feature vector of the entire UAV trajectory cluster. ;
[0077] S3.5: The time-series feature set extracted by the bidirectional GRU network and the population structure feature extracted by the graph neural network are fused through the fusion layer to obtain the fused features. This leads to the fused feature matrix. ;
[0078] S3.6: Use graph pooling to aggregate the fused features of all nodes into a global feature of a graph structure;
[0079] First, calculate the attention score for each node. The formula for calculating attention score is as follows:
[0080] (twenty one)
[0081] in, It is a weight matrix. It is a bias term;
[0082] Then, the attention scores are used to perform a weighted summation of the fused features of the nodes to generate the global features of the graph structure. The specific formula is as follows:
[0083] (twenty two)
[0084] S3.7: Global features of the graph structure The input is fed into a fully connected layer, where a softmax function maps the probability distribution of each intent, ultimately yielding the intent prediction result for the drone trajectory cluster. The specific formula is as follows:
[0085] (twenty three)
[0086] in, It is the predicted probability distribution. The drone trajectory cluster representing the input belongs to the intent. The probability, It is the total number of intentions. and These are the weight matrix and bias term of the fully connected layer, respectively.
[0087] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0088] This invention preprocesses 3D trajectories to ensure their consistency in time and space, improving data stability and quality, reducing noise interference, and making subsequent intent recognition more accurate and reliable. Through an improved HDBSCAN algorithm, this invention can effectively identify trajectory direction information and dynamically adjust the clustering structure at different density levels, making UAV intent recognition more accurate in different group sizes and complex task scenarios. Combined with deep learning networks, this invention can better capture the temporal dependencies of 3D trajectories, improving the ability to analyze complex behavioral patterns during intent recognition, thereby enhancing the accuracy and reliability of the recognition results. Especially in complex scenarios such as multi-UAV collaborative operations, it achieves more effective recognition of UAV swarm intent, significantly improving intent recognition performance compared to existing technologies. Attached Figure Description
[0089] Figure 1 This is a flowchart of an intent recognition method based on three-dimensional trajectory analysis of a drone swarm, as described in an embodiment of the present invention.
[0090] Figure 2 This is a structural diagram of the deep learning intent recognition network in an embodiment of the present invention;
[0091] Figure 3 This is a schematic diagram of the trajectory intention encoding of a drone swarm in an embodiment of the present invention. Detailed Implementation
[0092] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0093] like Figure 1 As shown, an intent recognition method based on 3D trajectory analysis of a drone swarm includes the following steps:
[0094] S1: Obtain the 3D trajectory dataset of the UAV cluster and preprocess the 3D trajectories to obtain the preprocessed 3D trajectory dataset; the 3D trajectory dataset of the UAV cluster includes the 3D trajectories of multiple UAVs, and each UAV's 3D trajectory includes several time steps and the 3D spatial coordinates of the trajectory points at that time step;
[0095] S1.1: Obtain the 3D trajectories of drones in the drone cluster, obtain the 3D trajectory dataset of the drone cluster, and filter the obtained 3D trajectories, deleting incomplete and abnormal 3D trajectories to ensure the validity of the 3D trajectories.
[0096] drones The three-dimensional trajectory is:
[0097] (1)
[0098] in, Indicates drone The three-dimensional trajectory This represents the time step, and 'i' represents the number of the trajectory point. Indicates the drone at time step The three-dimensional spatial coordinates of the trajectory points below, Indicates the number of trajectory points;
[0099] Set the minimum threshold for the length of the three-dimensional trajectory as follows: If the length of the three-dimensional trajectory If the three-dimensional trajectory is incomplete and lacks analytical value, it is deleted; the average velocity of the three-dimensional trajectory is calculated. If the average velocity of a certain three-dimensional trajectory exceeds a reasonable range If the three-dimensional trajectory is incorrect or abnormal, then it is considered to contain errors or anomalies. If so, then delete the 3D trajectory;
[0100] The formula for calculating average speed is:
[0101] (2)
[0102] in, This represents the distance between two adjacent points in a 3D trajectory. Indicates and Adjacent time steps;
[0103] S1.2: Use cubic spline interpolation to smooth the three-dimensional trajectory in order to further eliminate noise and repair missing data, while unifying the three-dimensional trajectory to the same time dimension;
[0104] Specifically, this involves constructing a cubic spline function based on the selected 3D trajectory. ,in The first step is used to interpolate missing trajectory points, achieving smoothing of the 3D trajectory; subsequently, a uniform resampling time step is applied. Resampling is performed to generate trajectory points at the same time step;
[0105] The resampled trajectory points are represented as follows:
[0106] (3)
[0107] in, This represents the trajectory points after resampling. and These are the start and end times of the three-dimensional trajectory, respectively.
[0108] S1.3: To facilitate subsequent analysis and modeling, all three-dimensional trajectories are standardized to have the same dimensions and range;
[0109] S1.4: To analyze the trajectory intent of a drone swarm more comprehensively and accurately, it is necessary to map the 3D trajectories collected from different small areas onto the same large map through coordinate transformation. This involves determining the 3D spatial coordinates of each trajectory point within the 3D trajectory. This is mapped to the coordinate system of the large map using the following matrix transformation formula;
[0110] (4)
[0111] in, For the transformation matrix, It is an offset vector. These are the three-dimensional spatial coordinates of the trajectory points mapped onto the large map;
[0112] S2: To achieve preliminary intent classification, improve the processing capability of 3D trajectories of drone swarms of different sizes, enhance the generalization of the model, and prepare for subsequent intent recognition, cluster analysis is performed on the preprocessed 3D trajectories to divide the 3D trajectories of drones with similar behavioral characteristics into different drone trajectory clusters. The 3D trajectories in the drone trajectory clusters represent the similar motion patterns of the drone swarms under specific conditions.
[0113] S2.1: Extract trajectory points with large changes in direction from the three-dimensional trajectory by calculating the changes in angle;
[0114] Set the three-dimensional trajectory by The trajectory consists of several trajectory points, and the three-dimensional spatial coordinates of the trajectory points in the three-dimensional trajectory are represented as follows: ;
[0115] First, for every two adjacent trajectory points on the 3D trajectory, construct a vector pointing from one trajectory point to the adjacent trajectory point. :
[0116] (5)
[0117] in, The three-dimensional spatial coordinates of adjacent trajectory points;
[0118] For two adjacent vectors and The angle between them It can be calculated using the dot product formula for vectors:
[0119] (6)
[0120] in, Represents the dot product of two vectors. and It is the magnitude of the vector, and the formula is:
[0121] (7)
[0122] For ease of comparison, the difference in motion direction between any two adjacent trajectory points on the 3D trajectory is calculated using the following formula:
[0123] (8)
[0124] in, This indicates the difference in the direction of motion between two adjacent trajectory points. The value range is from 0 to 2. A larger value indicates a greater change in the direction of the 3D trajectory. The threshold should be set according to specific requirements. ,when At that time, the trajectory point is recorded, thereby extracting the trajectory points with large changes in direction in the three-dimensional trajectory;
[0125] S2.2: Calculate the similarity measure between three-dimensional trajectories using the starting trajectory point, ending trajectory point, and extracted trajectory points with large direction changes in the three-dimensional trajectory;
[0126] The three-dimensional trajectory similarity measurement method proposed in this invention is a comprehensive trajectory similarity measurement method constructed from three-dimensional trajectory shape similarity (TSS) and three-dimensional trajectory direction similarity (TDS), which comprehensively measures the degree of similarity between two trajectories from both spatial shape and motion direction aspects.
[0127] To facilitate calculation, this invention uses the starting trajectory point, ending trajectory point, and trajectory points with significant changes in trajectory direction extracted in S2.1 to calculate the similarity between the three-dimensional trajectories. The specific two three-dimensional trajectories are expressed as follows:
[0128] (9)
[0129] (10)
[0130] in, and There are two three-dimensional trajectories. and Representing three-dimensional trajectories respectively and three-dimensional trajectory Trajectory points in; j are the numbers of the trajectory points. It is a three-dimensional trajectory The total number of starting trajectory points, ending trajectory points, and extracted trajectory points with significant changes in trajectory direction; It is a three-dimensional trajectory The total number of starting trajectory points, ending trajectory points, and extracted trajectory points with significant changes in trajectory direction;
[0131] First, calculate the three-dimensional trajectory separately. From the starting point, ending point, and points with significant direction changes on the trajectory to the three-dimensional trajectory The distance to the nearest trajectory point (called the Nearest Neighbor Distance Points, NNP) is then used to calculate the 3D trajectory. From the starting point, ending point, and points with significant direction changes on the trajectory to the three-dimensional trajectory The average distance to the nearest trajectory point is used to obtain the 3D trajectory using the same method. From the starting point, ending point, and points with significant direction changes on the trajectory to the three-dimensional trajectory The mean distance to the nearest trajectory point is taken, and then the minimum of these two values is taken as the 3D trajectory. and three-dimensional trajectory The similarity of the three-dimensional trajectory shapes between them is calculated using the following formula:
[0132] (11)
[0133] in, Representing a three-dimensional trajectory and three-dimensional trajectory The similarity of the three-dimensional trajectory shapes between them; Representing a three-dimensional trajectory trajectory points on To three-dimensional trajectory Distance to the nearest trajectory point Representing a three-dimensional trajectory trajectory points on To three-dimensional trajectory Distance to the nearest trajectory point;
[0134] Secondly, for three-dimensional trajectories Calculate the relationship between the starting trajectory point, ending trajectory point, and trajectory points with large direction changes on the trajectory and the three-dimensional trajectory. The mean of the differences in the motion directions of the nearest trajectory points is used to similarly calculate the three-dimensional trajectory. The starting point, ending point, and points with significant changes in direction of the trajectory are compared with the three-dimensional trajectory. The mean of the differences in the motion directions of the nearest trajectory points is taken as the minimum value of the differences between the two values, which is taken as the 3D trajectory direction similarity, representing the difference in the motion directions of the 3D trajectory. The formula for calculating the 3D trajectory direction similarity is as follows:
[0135] (12)
[0136] in, Representing a three-dimensional trajectory and three-dimensional trajectory The similarity of the three-dimensional trajectory directions between them; Representing a three-dimensional trajectory Upper trajectory point and three-dimensional trajectory Upper trajectory point The angle between the directions of motion, take The value is used to characterize the difference in the direction of motion;
[0137] Finally, considering that the two key features of 3D trajectory shape similarity and 3D trajectory direction similarity have equally important impacts on 3D trajectory similarity, and in order to achieve flexible adjustments in different scenarios, the 3D trajectory shape similarity and 3D trajectory direction similarity are normalized and then combined to comprehensively analyze their impact on 3D trajectory similarity, ensuring more accurate 3D trajectory comparison. The formula for calculating the similarity measure between 3D trajectories is as follows:
[0138] (13)
[0139] in, This represents the maximum value of the similarity between the three-dimensional trajectory shapes. This is a similarity measure between three-dimensional trajectories, with values ranging from... A value closer to 1 indicates greater similarity between the two 3D trajectories; conversely, a smaller value indicates greater difference between the trajectories. When the two 3D trajectories move in the same direction, the similarity is primarily determined by the shape similarity of the trajectories; however, when the two trajectories move in completely opposite directions... ,lead to This indicates that the differences between the three-dimensional trajectories are extremely large;
[0140] S2.3: Based on the similarity measure between three-dimensional trajectories, the improved HDBSCAN (HierarchicalDensity-Based Spatial Clustering of Applications with Noise) algorithm is used to perform cluster analysis on the three-dimensional trajectories of the preprocessed UAV swarm, and obtain several UAV trajectory clusters with similar behavioral patterns, effectively revealing the cooperative combat characteristics of the UAV swarm and providing an important basis for subsequent intent recognition.
[0141] The improved HDBSCAN algorithm replaces the core distance in the original HDBSCAN algorithm with the trajectory core distance. The original HDBSCAN algorithm's mutual reachability distance is improved to a trajectory mutual reachability distance. ;
[0142] The distance of the trajectory core The calculation method is as follows: calculate the custom distance function value between every two 3D trajectories. For a 3D trajectory... The custom distance function values between the trajectories and other 3D trajectories are arranged in ascending order, and the k-th custom distance function value is taken as the core distance of the trajectory. The default value of k is consistent with the minimum cluster size (min_cluster_size) parameter;
[0143] The custom distance function is defined based on the similarity metric between 3D trajectories, such that the smaller the custom distance, the more similar the 3D trajectories are. The formula for calculating the custom distance function is as follows:
[0144] (14)
[0145] in, For three-dimensional trajectory and three-dimensional trajectory Custom distance function values between them;
[0146] The reachability metric distance between the trajectories is used to disperse low-density trajectories (those with high trajectory core distances), as shown in the following formula:
[0147] (15)
[0148] in, This represents the distance between two 3D trajectories that are mutually reachable.
[0149] S3: Input each drone trajectory cluster into a pre-trained deep learning intent recognition network to perform intent recognition and obtain the intent prediction result for each drone trajectory cluster;
[0150] The intents of targets represented by different UAV trajectory clusters vary significantly. Therefore, a deep learning-based trajectory intent recognition method is proposed. This method inputs the 3D trajectories from different UAV trajectory clusters into the network for recognition and outputs the final intent of the UAV trajectory cluster, such as... Figure 2 and Figure 3 As shown, the intentions include attack, retreat, electronic jamming, surveillance, reconnaissance, feint, and transportation, with the specific steps as follows:
[0151] S3.1: Clustering a specific UAV trajectory cluster All the 3D trajectories are input into the input layer, where, , The number of 3D trajectories in the UAV trajectory cluster. The time step (length of the trajectory sequence) is represented by 4, indicating that each trajectory point includes both a time step and three-dimensional spatial coordinates.
[0152] S3.2: Construct a graph structure based on the 3D trajectories in the UAV trajectory cluster;
[0153] S3.2.1: Treat each 3D trajectory in the UAV trajectory cluster as a node and initialize the node features. Each node corresponds to one UAV. Initialize as a feature vector And construct the initial feature matrix. ;
[0154] The feature vector Represents a node The initial features, where It is a node The mean of all three-dimensional spatial coordinates in the corresponding three-dimensional trajectory of the UAV;
[0155] S3.2.2: Construct a weighted adjacency matrix based on the similarity measure between the three-dimensional trajectories of the UAVs. ;
[0156] Specifically, it involves: a similarity measurement based on three-dimensional trajectories. Construct a similarity matrix , the elements Set similarity threshold ,like Not less than the threshold Then the node and nodes There are edges between them, weighted adjacency matrix elements in This is a similarity measure between 3D trajectories; if there are no edges, then the element... =0;
[0157] (16)
[0158] S3.2.3: To ensure that each node can include its own information during feature aggregation, in the weighted adjacency matrix... By adding self-loops, each node can aggregate its own information to obtain a weighted adjacency matrix with self-loops added.
[0159] The adjacency matrix after adding the self-loop is: ,in It is the identity matrix;
[0160] S3.2.4: Based on the initial characteristic matrix and the adjacency matrix after adding self-loops, the graph structure is obtained;
[0161] S3.3: In order for the network to extract the group behavior features of UAV trajectory clusters, a graph structure is input into a two-layer graph neural network (GNN). The two-layer graph neural network realizes feature sharing within the UAV trajectory cluster, that is, by aggregating the feature vector of each node with the feature vector of its neighboring nodes to generate group structure features.
[0162] In GCN (Graph Convolutional Network), the feature vector update of each layer's nodes follows the formula:
[0163] (17)
[0164] in, It is the first The feature matrix of the layer, It is the first The feature matrix of the layer, where each row represents the feature vector of a node. For the initial layer, This is the initial feature matrix; yes The degree matrix, whose diagonal elements , These are the elements in the weighted adjacency matrix after adding self-loops; It is the first The weight matrix of the layer needs to be learned through training; It is a non-linear activation function, ReLU.
[0165] In a two-layer GCN, the feature update formulas for the first and second layers are as follows:
[0166] (18)
[0167] (19)
[0168] Through the first layer update, the node feature matrix The first layer already includes neighborhood information, integrating the features of each node and its neighbors; the second layer further aggregates the neighborhood information to obtain the output feature matrix of the two-layer GCN. , as a characteristic of group structure;
[0169] S3.4: Input the 3D trajectory in the UAV trajectory cluster into the bidirectional GRU network. Bi-GRU can simultaneously capture the forward and backward dependencies of the 3D trajectory to better capture the global 3D trajectory change features, thereby effectively learning the change pattern of the 3D trajectory over time, extracting the time series feature sequence set of each 3D trajectory in the UAV trajectory cluster, and then obtaining the time feature vector of the UAV trajectory cluster.
[0170] (20)
[0171] in, Indicates time step Drone The three-dimensional trajectory is in the output of the bidirectional GRU. In time step The input 3D trajectory, It is the output of the forward GRU. It is the output of the backward GRU;
[0172] Finally, based on the output of the bidirectional GRU (Bi-GRU) at each time step in the three-dimensional trajectory, the UAV is obtained. Time series feature set , The number of time steps is used to obtain the temporal feature vector of the entire UAV trajectory cluster. ;
[0173] S3.5: The fusion layer fuses the time-series feature set extracted by the bidirectional GRU network and the group structure features extracted by the GCN to reconstruct a feature representation containing richer information, which facilitates further analysis of the intent of the UAV swarm and yields the fused features. This leads to the fused feature matrix. ;
[0174] S3.6: Use graph pooling to aggregate the fused features of all nodes into a global feature graph structure;
[0175] First, calculate the attention score for each node. The formula for calculating attention score is as follows:
[0176] (twenty one)
[0177] in, It is a weight matrix. It is a bias term, and softmax is used to standardize the attention scores of all nodes;
[0178] Then, the attention scores are used to perform a weighted summation of the fused features of the nodes to generate the global features of the graph structure. The specific formula is as follows:
[0179] (twenty two)
[0180] S3.7: Global features of the graph structure The input is fed into a fully connected layer, where a softmax function maps the probability distribution of each intent, ultimately yielding the intent prediction result for the drone trajectory cluster. The specific formula is as follows:
[0181] (twenty three)
[0182] in, It is the predicted probability distribution, containing the probabilities of each intention, where This represents the intent of the input drone trajectory cluster. The probability, It is the total number of intentions. and These are the weight matrix and bias term of the fully connected layer, respectively. Through softmax, the network outputs the intent classification result corresponding to the drone trajectory cluster, and finally identifies the target's behavioral intent.
[0183] The training method for the pre-trained deep learning intent recognition network in this embodiment is as follows:
[0184] To identify the intent of drone trajectory clusters, the individual drone trajectory clusters within the swarm must first be labeled. Specifically, for each drone trajectory cluster, an intent label is manually assigned, with intent categories including but not limited to attack, retreat, electronic jamming, surveillance, reconnaissance, feint, and transportation. After labeling, the dataset is divided into training, validation, and test sets, typically with 70% of the data used for training, 15% for validation, and 15% for testing.
[0185] During training, forward propagation is first performed on each batch of training samples. The drone trajectory cluster is then input into the model, calculated layer by layer, and finally outputs the predicted intent of that trajectory cluster. To evaluate how well the model's predictions match the true labels, the cross-entropy loss function is used to calculate the loss, as shown in the following formula:
[0186] (twenty four)
[0187] in, Show the first The true label of the class, It is the model's predicted probability for that category.
[0188] Next, the backpropagation algorithm is used to calculate the gradient of the loss with respect to the parameters of each layer, and the model parameters are updated using the Adam optimizer to gradually minimize the loss. After each training epoch, the model performance is evaluated using validation set data, and the validation loss and accuracy are recorded. When the validation loss no longer decreases significantly over several consecutive training epochs, an early stopping strategy is applied to terminate training to prevent overfitting.
[0189] After training, the model is evaluated on the test set, and metrics such as accuracy, precision, recall, and F1-score are calculated to comprehensively assess the model's generalization performance. Based on the performance on the validation set, hyperparameters (such as learning rate and regularization coefficient) are adjusted as necessary, and the model is retrained. Finally, the optimal model is selected for actual deployment.
[0190] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. An intent recognition method based on three-dimensional trajectory analysis of UAV swarms, characterized in that, Includes the following steps: S1: Obtain the 3D trajectory dataset of the UAV cluster and preprocess the 3D trajectories to obtain the preprocessed 3D trajectories; the 3D trajectory dataset of the UAV cluster includes the 3D trajectories of multiple UAVs, and each UAV's 3D trajectory includes several time steps and the 3D spatial coordinates of the trajectory points at that time step; S2: Perform cluster analysis on the preprocessed 3D trajectory to divide it into different UAV trajectory clusters; S2 specifically includes: S2.1: Extract trajectory points with large changes in direction from the three-dimensional trajectory by calculating the changes in angle; S2.2: Calculate the similarity measure between three-dimensional trajectories using the starting trajectory point, ending trajectory point, and extracted trajectory points with large direction changes in the three-dimensional trajectory; S2.2 specifically includes: The two three-dimensional trajectories are expressed as follows: in, and There are two three-dimensional trajectories. and Representing three-dimensional trajectories respectively and three-dimensional trajectory The trajectory points in the diagram; i and j are the numbers of the trajectory points. It is a three-dimensional trajectory The total number of starting trajectory points, ending trajectory points, and extracted trajectory points with significant changes in trajectory direction; It is a three-dimensional trajectory The total number of starting trajectory points, ending trajectory points, and extracted trajectory points with significant changes in trajectory direction; First, calculate the three-dimensional trajectory separately. From the starting point, ending point, and points with significant direction changes on the trajectory to the three-dimensional trajectory Calculate the distance to the nearest trajectory point, and then calculate the 3D trajectory. From the starting point, ending point, and points with significant direction changes on the trajectory to the three-dimensional trajectory The average distance to the nearest trajectory point is used to obtain the 3D trajectory using the same method. From the starting point, ending point, and points with significant direction changes on the trajectory to the three-dimensional trajectory The mean distance to the nearest trajectory point is taken, and then the minimum of these two values is taken as the 3D trajectory. and three-dimensional trajectory The similarity of the three-dimensional trajectory shapes between them is calculated using the following formula: in, Representing a three-dimensional trajectory and three-dimensional trajectory The similarity of the three-dimensional trajectory shapes between them; Representing a three-dimensional trajectory trajectory points on To three-dimensional trajectory Distance to the nearest trajectory point Representing a three-dimensional trajectory trajectory points on To three-dimensional trajectory Distance to the nearest trajectory point; Secondly, for three-dimensional trajectories Calculate the relationship between the starting trajectory point, ending trajectory point, and trajectory points with large direction changes on the trajectory and the three-dimensional trajectory. The mean of the differences in the motion directions of the nearest trajectory points is used to similarly calculate the three-dimensional trajectory. The starting point, ending point, and points with significant changes in direction of the trajectory are compared with the three-dimensional trajectory. The mean of the differences in the motion directions of the nearest trajectory points is taken as the minimum of these two values, and the formula for calculating the three-dimensional trajectory direction similarity is as follows: in, Representing a three-dimensional trajectory and three-dimensional trajectory The similarity of the three-dimensional trajectory directions between them; Representing a three-dimensional trajectory Upper trajectory point and three-dimensional trajectory Upper trajectory point The angle between the directions of motion, take The value is used to characterize the difference in the direction of motion; Finally, the similarity between the three-dimensional trajectory shape and the three-dimensional trajectory direction is normalized and then combined to calculate the similarity measure between the three-dimensional trajectories. The calculation formula is as follows: in, This represents the maximum value of the similarity between the three-dimensional trajectory shapes. This is a similarity measure between three-dimensional trajectories, with values ranging from... ; S2.3: Based on the similarity measure between the three-dimensional trajectories, the improved HDBSCAN algorithm is used to perform cluster analysis on the preprocessed three-dimensional trajectories to obtain several UAV trajectory clusters; The improved HDBSCAN algorithm replaces the core distance in the original HDBSCAN algorithm with the trajectory core distance. The original HDBSCAN algorithm's mutual reachability metric distance is improved to a trajectory mutual reachability metric distance. ; The distance of the trajectory core The calculation method is as follows: calculate the custom distance function value between every two 3D trajectories. For a 3D trajectory... The custom distance function values between the trajectories and other 3D trajectories are arranged in ascending order, and the k-th custom distance function value is taken as the core distance of the trajectory. ; The formula for calculating the custom distance function is as follows: in, For three-dimensional trajectory and three-dimensional trajectory Custom distance function values between them; The distance between the mutually reachable trajectories is calculated using the following formula: in, This represents the distance between two 3D trajectories that are mutually reachable. S3: Input each drone trajectory cluster into a pre-trained deep learning intent recognition network for intent recognition, and obtain the intent prediction result for each drone trajectory cluster.
2. The intention recognition method based on three-dimensional trajectory analysis of UAV swarms according to claim 1, characterized in that, S1 specifically includes: S1.1: Obtain the 3D trajectories of drones in the drone cluster, obtain the 3D trajectory dataset of the drone cluster, and filter the obtained 3D trajectories, deleting incomplete and abnormal 3D trajectories. drones The three-dimensional trajectory is: in, Indicates drone The three-dimensional trajectory This represents the time step, and 'i' represents the number of the trajectory point. Indicates the drone at time step The three-dimensional spatial coordinates of the trajectory points below, Indicates the number of trajectory points; Set the minimum threshold for the length of the three-dimensional trajectory as follows: If the length of the three-dimensional trajectory If the three-dimensional trajectory is incomplete, it is considered incomplete and deleted; calculate the average velocity of the three-dimensional trajectory. If the average velocity of a certain three-dimensional trajectory exceeds a reasonable range If the three-dimensional trajectory is incorrect or abnormal, then it is considered to contain errors or anomalies. If so, then delete the 3D trajectory; S1.2: Use cubic spline interpolation to smooth the three-dimensional trajectory and unify the three-dimensional trajectory to the same time dimension; Specifically, this involves constructing a cubic spline function based on the selected 3D trajectory. ,in The first step is used to interpolate missing trajectory points; then, a uniform resampling time step is applied. Resampling is performed to generate trajectory points at the same time step; The resampled trajectory points are represented as follows: in, This represents the trajectory points after resampling. and These are the start and end times of the three-dimensional trajectory, respectively. S1.3: Standardize all three-dimensional trajectories to make them have the same dimensions and range; S1.4: Map the 3D trajectory to the same large map through coordinate transformation; For the three-dimensional spatial coordinates of each trajectory point in the three-dimensional trajectory This is mapped to the coordinate system of the large map using the following matrix transformation formula: in, For the transformation matrix, It is an offset vector. These are the three-dimensional spatial coordinates of the trajectory points mapped onto the large map.
3. The intention recognition method based on three-dimensional trajectory analysis of UAV swarms according to claim 2, characterized in that, S2.1 specifically includes: Set the three-dimensional trajectory by The trajectory consists of several trajectory points, and the three-dimensional spatial coordinates of the trajectory points in the three-dimensional trajectory are represented as follows: ; First, for every two adjacent trajectory points on the 3D trajectory, construct a vector pointing from one trajectory point to the adjacent trajectory point. : in, The three-dimensional spatial coordinates of adjacent trajectory points; For two adjacent vectors and The angle between them It can be calculated using the dot product formula for vectors: in, Represents the dot product of two vectors. and It is the magnitude of the vector, and the formula is: For every two adjacent trajectory points on the three-dimensional trajectory, calculate the difference in their motion direction. The specific calculation formula is as follows: in, This indicates the difference in the direction of motion between two adjacent trajectory points. The value range is from 0 to 2. A larger value indicates a greater change in the direction of the 3D trajectory. The threshold should be set according to specific requirements. ,when At that time, the trajectory point is recorded, thereby extracting the trajectory points with large changes in direction in the three-dimensional trajectory.
4. The intent recognition method based on three-dimensional trajectory analysis of UAV swarms according to claim 3, characterized in that, S3 specifically includes: S3.1: Clustering a specific UAV trajectory cluster All the 3D trajectories are input into the input layer, where, , The number of 3D trajectories in the UAV trajectory cluster. For time step; S3.2: Construct a graph structure based on the 3D trajectories in the UAV trajectory cluster; S3.3: Input the graph structure into a two-layer graph neural network, and realize feature sharing within the UAV trajectory cluster through the two-layer graph neural network. That is, the feature vector of each node is aggregated with the feature vectors of its neighboring nodes to generate group structure features. In each layer of the graph convolutional network, the node feature vector update follows the formula: in, It is the first The feature matrix of the layer, It is the first The feature matrix of the layer, where each row represents the feature vector of a node. For the initial layer, This is the initial feature matrix; yes The degree matrix, whose diagonal elements , These are the elements in the weighted adjacency matrix after adding self-loops; It is the first Layer weight matrix; It uses the non-linear activation function ReLU to obtain the output feature matrix of the two-layer graph neural network. , as a characteristic of group structure; S3.4: Input the three-dimensional trajectory in the UAV trajectory cluster into the bidirectional GRU network, extract the time series feature sequence set of each three-dimensional trajectory in the UAV trajectory cluster, and then obtain the time feature vector of the UAV trajectory cluster; in, Indicates time step Drone The three-dimensional trajectory is in the output of the bidirectional GRU. In time step The input 3D trajectory, It is the output of the forward GRU. It is the output of the backward GRU; Finally, based on the output of the bidirectional GRU (Bi-GRU) at each time step in the three-dimensional trajectory, the UAV is obtained. Time series feature set , The number of time steps is used to obtain the temporal feature vector of the entire UAV trajectory cluster. ; S3.5: The time-series feature set extracted by the bidirectional GRU network and the population structure feature extracted by the graph neural network are fused through the fusion layer to obtain the fused features. This leads to the fused feature matrix. ; S3.6: Use graph pooling to aggregate the fused features of all nodes into a global feature of a graph structure; First, calculate the attention score for each node. The formula for calculating attention score is as follows: in, It is a weight matrix. It is a bias term; Then, the attention scores are used to perform a weighted summation of the fused features of the nodes to generate the global features of the graph structure. The specific formula is as follows: S3.7: Global features of the graph structure The input is fed into a fully connected layer, where a softmax function maps the probability distribution of each intent, ultimately yielding the intent prediction result for the drone trajectory cluster. The specific formula is as follows: in, It is the predicted probability distribution. The drone trajectory cluster representing the input belongs to the intent. The probability, It is the total number of intentions. and These are the weight matrix and bias term of the fully connected layer, respectively.
5. The intention recognition method based on three-dimensional trajectory analysis of UAV swarms according to claim 4, characterized in that, S3.2 specifically includes: S3.2.1: Treat each UAV corresponding to a 3D trajectory in the UAV trajectory cluster as a node, and initialize the node features. Initialize as a feature vector And construct the initial feature matrix. ; The feature vector Represents a node The initial features, of which It is a node The mean of all three-dimensional spatial coordinates in the corresponding three-dimensional trajectory of the UAV; S3.2.2: Construct a weighted adjacency matrix based on the similarity measure between the three-dimensional trajectories of the UAVs. ; Specifically, it involves: a similarity measurement based on three-dimensional trajectories. Construct a similarity matrix , the elements Set similarity threshold ,like Not less than the threshold Then the node and nodes There are edges between them, weighted adjacency matrix elements in This is a similarity measure between 3D trajectories; if there are no edges, then the element... =0; S3.2.3: In the weighted adjacency matrix Adding self-loops to the matrix yields a weighted adjacency matrix with added self-loops. The adjacency matrix after adding the self-loop is: ,in It is the identity matrix; S3.2.4: Based on the initial characteristic matrix and the adjacency matrix after adding self-loops, the graph structure is obtained.