Target behavior rule mining method based on depth map clustering
By using a deep graph clustering method, and leveraging an autoencoder with an attention mechanism and a self-trained graph neural network, the difficulty of mining and analyzing high-dimensional target data in traditional methods is solved, enabling an efficient and accurate description of target behavior patterns.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NAVAL AVIATION UNIV
- Filing Date
- 2024-11-25
- Publication Date
- 2026-06-09
Smart Images
Figure CN119622374B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to clustering analysis techniques in data mining and high-level fusion techniques in information fusion, belonging to the fields of pattern recognition and intelligent intelligence processing. Background Technology
[0002] In complex scenarios of early warning and surveillance, characterized by high-intensity confrontation, multiple targets, and high mobility, the monitored area is characterized by dense, maneuverable, and camouflaged maritime and air targets. Target perception methods are diverse, including radar, electro-optical, sonar, AIS, and ADS-B. Target data in complex scenarios includes real-time high-dimensional target status data acquired by multiple sensors, such as target attributes, type, category, model, longitude, latitude, altitude, speed, and heading; historical target status data; historical activity patterns; sensor deployment; typical tasks; procedural rules; and control areas. When analyzing target behavior, target-related data can be represented as a high-dimensional trajectory map. Nodes in the map contain high-dimensional feature information about the target at a given moment, such as attributes, type, category, model, longitude, latitude, altitude, speed, heading, historical activity patterns, and current task. The edges connecting nodes contain dynamic information such as the relationship between the same target at different times, the relationship between different targets at the same time, and the relationship between different targets at different times.
[0003] Clustering target data can enable the analysis and discovery of patterns in target behavior. However, traditional clustering methods rely on the construction of similarity metrics, which suffer from problems such as strong data dependencies, high computational complexity, and inaccurate metric descriptions when dealing with complex, high-dimensional target data, thus limiting their effectiveness in uncovering patterns in target behavior. Summary of the Invention
[0004] This invention addresses the limitations of existing methods for mining target activity patterns by providing a method for mining target behavior patterns based on deep graph clustering. A deep autoencoder with an attention mechanism is designed to construct a self-trained graph neural network model. An unsupervised learning mechanism is built through a bidirectional mutual supervision mode, enabling the mining and analysis of activity patterns of targets performing diverse tasks such as patrolling, transportation, and navigation.
[0005] The method for mining target behavior patterns based on depth map clustering of the present invention is characterized by including the following steps:
[0006] Step 1: Set the target's attributes and type tags;
[0007] Step 2: Represent the target data in the form of a spatiotemporal diagram. ;
[0008] Step 3: Design a spatiotemporal graph autoencoder with an attention mechanism. Learn node representations by aggregating adjacency matrix information and reconstruct the spatiotemporal graph network structure by calculating the inner product of node pairs.
[0009] Step 4: Construct a self-trained graph neural network model, aggregate the dimensionality-reduced nearest neighbor target information, and mine the structural information between targets;
[0010] Step 5: Design a dual self-supervised module consisting of a graph convolutional neural network and a deep neural network. Construct an unsupervised learning mechanism for the model through a bidirectional mutual supervision mode to obtain the cluster labels of the target node set.
[0011] Step 6: Set target behavior pattern tags;
[0012] Step 7: Visualize the patterns of the target behavior.
[0013] Preferably, step 1 uses target attributes, types, and cluster number tags to form target behavior pattern tags, thereby representing the target's regular behavior. In the field of early warning and surveillance, the target attributes are divided into friendly, enemy, and allied, represented by tags 1, 2, and 3 respectively. The target types are divided into military aircraft, civilian aircraft, warships, and civilian ships, represented by tags 1, 2, 3, and 4 respectively.
[0014] Preferably, step 2 represents the target data in the form of a spatiotemporal diagram. ,in, It is a set of nodes. and They represent Given the edge set and adjacency matrix at time t, the spatiotemporal graph is a directed weighted heterogeneous dynamic graph.
[0015] Preferably, the spatiotemporal graph autoencoder with attention mechanism described in step 3 is used to update the aggregated neighbor information model of the node representation, as shown in the following equation:
[0016] ;
[0017] in, , These are the nodes before and after aggregating neighbor information. The expression, Representative node The neighborhood group, This represents node pairs. Attention weights between them;
[0018] With node representations, the original network structure is reconstructed by calculating the inner product of node pairs, thereby achieving unsupervised node representation learning.
[0019] ;
[0020] in, This can be understood as node pairs The probability that an edge exists between them.
[0021] Preferably, when constructing the self-trained graphical neural network model in step 4, the cluster center is set as... , then the node Probability of belonging to a certain category As shown in the following formula:
[0022] ;
[0023] in, It can be viewed as a distribution of node allocation. To determine the number of nodes and to incorporate clustering information to achieve cluster-oriented node representation, making each node closer to its corresponding cluster center and minimizing intra-cluster distance and maximizing inter-cluster distance, the following target distribution is defined:
[0024] ;
[0025] Finally, by calculating the difference between the two distributions... KL Divergence is used to achieve mutual constraints, which is also known as self-training.
[0026] ;
[0027] This is the loss function for the model;
[0028] node The cluster to which a belongs can be calculated using the following formula:
[0029] .
[0030] Preferably, in step 5, when designing the dual self-supervised module of graph convolutional neural network and deep neural network, the following will be implemented: Supervise the graph convolutional neural network module by treating it as a label:
[0031] ;
[0032] distributed It can act as a bridge, constraining the representations learned by deep neural networks and graph convolutional neural networks. The complete loss function of the model is as follows:
[0033] ;
[0034] , These are the weighting coefficients;
[0035] node The formula for calculating cluster tags has been updated to:
[0036] .
[0037] Preferably, step 6 sets target behavior pattern labels. In the field of early warning and surveillance, the clusters obtained after clustering the multi-dimensional flight path data of enemy military aircraft are... Then the labels for the target behavior patterns can be set sequentially as 211, 213, ..., 21m.
[0038] Preferably, the specific method for visualizing the target pattern in step 7 is as follows:
[0039] Select each cluster The track with the most nearest neighbors is used as the characteristic track of the behavior pattern. When displaying the visualization, only the characteristic track is drawn, and the target behavior pattern label corresponding to the characteristic track is marked. The target heading is indicated by arrows, and the speed of the target is indicated by the length of the interval between adjacent track points.
[0040] The beneficial effects of this invention are as follows: This invention provides a target behavior pattern mining method based on depth graph clustering, which makes full use of the high-dimensional information of the target and the information on the relationship between the targets. It can effectively solve the problems of the target activity pattern mining based on traditional clustering methods relying on the construction of similarity measurement, and having strong data dependence, high computational complexity, and inaccurate measurement description when facing high-dimensional target data in complex scenarios. It can achieve efficient mining and analysis of target behavior patterns. Attached Figure Description
[0041] Figure 1 This is an overall flowchart of the target behavior pattern mining method based on depth map clustering of the present invention. Detailed Implementation
[0042] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0043] Step 1: Set target attribute and type labels. This method uses target attributes, types, and cluster number labels to create target behavior pattern labels, thereby representing the target's regular behavior. In the field of early warning and surveillance, target attributes are divided into friendly, enemy, and allied, represented by labels 1, 2, and 3 respectively. Target types are divided into military aircraft, civilian aircraft, warships, and civilian vessels, represented by labels 1, 2, 3, and 4 respectively.
[0044] Step 2: Represent the target-related data in the form of a spatiotemporal diagram. ,in, It is a set of nodes. and They represent Given the edge set and adjacency matrix at time t, the spatiotemporal graph is a directed weighted heterogeneous dynamic graph.
[0045] Step 3: Design a spatiotemporal graph autoencoder with an attention mechanism. This autoencoder learns node representations by aggregating adjacency matrix information and reconstructs the spatiotemporal graph network structure by calculating the inner product of node pairs. The aggregated neighbor information model used to update node representations is shown in the following equation:
[0046] ;
[0047] in, , These are the nodes before and after aggregating neighbor information. The expression, Representative node The neighborhood group, This represents node pairs. Attention weights between them.
[0048] With node representations, the original network structure can be reconstructed by calculating the inner product of node pairs, thereby achieving unsupervised node representation learning.
[0049] ;
[0050] in, This can be understood as node pairs The probability that an edge exists between them.
[0051] Step 4: Construct a self-trained graph neural network model, aggregate the dimensionality-reduced nearest neighbor object information, and mine the structural information between objects. Let the cluster center be... Then the probability that a node belongs to a certain category. As shown in the following formula:
[0052] ;
[0053] here, This can be viewed as the distribution of node allocation. Let be the number of nodes. To incorporate clustering information to achieve cluster-oriented node representation, making each node closer to its corresponding cluster center, thus minimizing intra-cluster distance and maximizing inter-cluster distance, the target distribution is defined as follows:
[0054] ;
[0055] Finally, by calculating the difference between the two distributions... KL Divergence is used to achieve mutual constraints, which is also known as self-training.
[0056] ;
[0057] This is the loss function for the model.
[0058] node The cluster to which a belongs can be calculated using the following formula:
[0059] ;
[0060] Step 5: Design a dual self-supervised module consisting of a graph convolutional neural network and a deep neural network. Construct an unsupervised learning mechanism for the model through a bidirectional mutual supervision mode to obtain cluster labels for the set of nodes with the same objective.
[0061] Will Supervise the graph convolutional neural network module by treating it as a label:
[0062] ;
[0063] distributed It can act as a bridge, constraining the representations learned by deep neural networks and graph convolutional neural networks. The complete loss function of the model is as follows:
[0064] ;
[0065] , These are the weighting coefficients.
[0066] node The formula for calculating cluster tags has been updated to:
[0067] ;
[0068] Step 6, set target behavior pattern labels. In the field of early warning and surveillance, the clusters obtained after clustering the multi-dimensional flight path data of enemy military aircraft are... Then the labels for the target behavior patterns can be set sequentially as 211, 213, ..., 21m.
[0069] Step 7: Visualize the target's behavioral patterns. To visually represent the target's behavioral patterns, we can select the track with the most nearest neighbors in each cluster as the feature track of the behavioral pattern. In the visualization, only the feature track is drawn, and the corresponding target behavioral pattern label is marked. Arrows indicate the target's heading, and the length of the interval between adjacent track points indicates the target's speed.
[0070] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.
Claims
1. A method for mining target behavior patterns based on depth map clustering, characterized in that, Includes the following steps: Step 1: Set the target's attributes and type tags; Step 2: Represent the target data in the form of a spatiotemporal diagram. ; Step 3: Design a spatiotemporal graph autoencoder with an attention mechanism. Learn node representations by aggregating adjacency matrix information and reconstruct the spatiotemporal graph network structure by calculating the inner product of node pairs. Step 4: Construct a self-trained graph neural network model, aggregate the dimensionality-reduced nearest neighbor target information, and mine the structural information between targets; Step 5: Design a dual self-supervised module consisting of a graph convolutional neural network and a deep neural network. Construct an unsupervised learning mechanism for the model through a bidirectional mutual supervision mode to obtain the cluster labels of the target node set. Step 6: Set target behavior pattern tags; Step 7: Visualize the patterns of the target behavior; The spatiotemporal graph autoencoder with attention mechanism described in step 3, used to update the aggregated neighbor information model of node representations, is shown in the following equation: ； in, , These are the nodes before and after aggregating neighbor information. The expression, Representative node The neighborhood group, This represents node pairs. Attention weights between them; With node representations, the original network structure is reconstructed by calculating the inner product of node pairs, thereby achieving unsupervised node representation learning. ； in, For node pairs The probability that an edge exists between them.
2. The target behavior pattern mining method based on depth graph clustering according to claim 1, characterized in that, When constructing the self-trained graphical neural network model in step 4, let the cluster center be... , then the node Probability of belonging to a certain category As shown in the following formula: ; in, It can be viewed as a distribution of node allocation. Let be the number of nodes. To incorporate clustering information and implement cluster-oriented node representation, making each node closer to its corresponding cluster center, thus minimizing intra-cluster distance and maximizing inter-cluster distance, the following target distribution is defined: ； Finally, by calculating the difference between the two distributions... KL Divergence is used to achieve mutual constraints, i.e., self-training. ； This is the loss function for the model; node The cluster to which a belongs is calculated using the following formula: .
3. The target behavior pattern mining method based on depth graph clustering according to claim 1, characterized in that, When designing the dual self-supervised module of graph convolutional neural network and deep neural network in step 5, Supervise the graph convolutional neural network module by treating it as a label: ； distributed It acts as a bridge, constraining the representations learned by the deep neural network and the graph convolutional neural network. The complete loss function of the model is as follows: ； , These are the weighting coefficients; node The formula for calculating cluster tags has been updated to: .
4. The target behavior pattern mining method based on depth graph clustering according to claim 1, characterized in that, The specific method for visualizing the target behavior patterns described in step 7 is as follows: The track with the most nearest neighbors in each cluster is selected as the feature track of the behavior pattern. In the visualization display, only the feature track is drawn, and the target behavior pattern label corresponding to the feature track is marked. The target heading is indicated by arrows, and the speed of the target is indicated by the length of the interval between adjacent track points.