Target behavior rule mining method based on depth map clustering

Abstract

The invention relates to a clustering analysis technology in data mining and a high-level fusion technology in information fusion, and belongs to the field of pattern recognition and intelligent information processing. Comprising the following steps: 1) setting attributes and type labels of targets; 2) representing the target data in a space-time diagram form; 3) designing a space-time diagram auto-encoder with an attention mechanism, learning node representation by aggregating adjacent matrix information, and reconstructing a space-time diagram network structure by calculating an inner product of a node pair; 4) constructing a self-training graph neural network model, and aggregating neighbor target information after dimension reduction; according to the method, the problems that a traditional clustering method is high in data dependence, high in calculation complexity, inaccurate in measurement description and the like can be solved, and efficient mining analysis of the target behavior law is achieved.

Description

Technical Field

[0001] The present invention relates to cluster analysis technology in data mining and high-level fusion technology in information fusion, and belongs to the field of pattern recognition and intelligent intelligence processing. Background Art

[0002] In complex scenarios of early warning and surveillance such as strong confrontation, multiple targets, and high mobility, the sea and air targets in the surveillance area are dense, mobile, and camouflaged, and the means of target perception are diverse, including radar, optoelectronics, sonar, AIS, ADS-B, etc. Complex scenario target data includes real-time high-dimensional target status data such as target attributes, types, categories, models, longitude, latitude, altitude, speed, and heading acquired by multiple sensors, 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 in the form of a high-dimensional track graph. The nodes in the graph contain high-dimensional feature information such as the attributes, types, categories, models, longitude, latitude, altitude, speed, heading, historical activity patterns, and current tasks of the target at a certain moment. 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] By clustering the target data, we can mine and analyze the target behavior patterns. The mining of target activity patterns based on traditional clustering methods relies on the construction of similarity metrics. When faced with complex scenes and high-dimensional target data, there are problems such as strong data dependence, high computational complexity, and inaccurate measurement descriptions, which have limitations in mining target behavior patterns. Summary of the invention

[0004] Aiming at the limitations of existing methods for mining target activity patterns, the present invention provides 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-training graph neural network model. The unsupervised learning mechanism of the model is constructed through a two-way mutual supervision mode to realize the mining and analysis of the activity patterns of the target when performing various tasks such as patrolling, transportation, and sailing.

[0005] The target behavior law mining method based on deep graph clustering of the present invention is special in that it 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 graph ; Step 3: Design a spatiotemporal graph autoencoder with an attention mechanism to 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: Build a self-training graph neural network model, aggregate the nearest neighbor target information after dimensionality reduction, and mine the structural information between targets; Step 5: Design a dual self-supervisory module of graph convolutional neural network and deep neural network, build an unsupervised learning mechanism of the model through a two-way mutual supervision mode, and obtain the cluster labels of the target node set; Step 6: Set the target behavior pattern label; Step 7: Visualize the target behavior patterns.

[0006] Preferably, in step 1, target behavior regularity labels are formed by setting target attributes, types and cluster number labels, thereby indicating the regular behavior of the target. In the field of early warning and monitoring, the attributes of the target are divided into our side, the enemy side and the friendly side, which are indicated by labels 1, 2 and 3 respectively, and the target types are divided into military aircraft, civil aircraft, warships and civilian ships, etc., which are indicated by labels 1, 2, 3 and 4 respectively; Preferably, the step 2 represents the target data in the form of a spatiotemporal graph. ,in, is a collection of nodes, and Respectively The edge set and adjacency matrix at each moment, the space-time graph is a directed weighted heterogeneous dynamic graph.

[0007] Preferably, the spatiotemporal graph autoencoder with an attention mechanism is designed in step 3, and the aggregated neighbor information model for updating the node representation is shown as follows: ; in, , They are the nodes before and after aggregating neighbor information. The expression, Representative Node The neighbor set of Represents the node pair The attention weight between After obtaining the node representation, the original network structure is reconstructed by calculating the inner product of the node pairs, thereby realizing unsupervised node representation learning: ; in, It can be understood as a node pair The probability that there is an edge between them.

[0008] Preferably, when constructing the self-training graph neural network model in step 4, the cluster center is set to , then the node The probability of belonging to a certain category , as shown below: ; in, As the distribution of node allocation, is the number of nodes. In order to introduce clustering information to realize clustering-oriented node representation, each node is closer to the corresponding cluster center, and the distance within the class is minimized and the distance between classes is maximized. Therefore, the following target distribution is defined: ; Finally, by calculating the difference between the two distributions KL Divergence is used to achieve mutual constraints, which is called self-training: ; is the loss function of the model; node The cluster in which can be calculated by the following formula: .

[0009] Preferably, when designing the dual self-supervision module of graph convolutional neural network and deep neural network in step 5, Considered as labels to supervise the graph convolutional neural network module: ; distributed It can play a bridging role, constraining the representations learned by the deep neural network and the representations learned by the graph convolutional neural network. The complete loss function of the model is as follows: ; , is the weight coefficient; node The calculation formula for cluster labels is updated as: .

[0010] Preferably, the step 6 sets the target behavior law label. In the field of early warning and monitoring, the clusters obtained after clustering the multi-dimensional track data of the enemy military aircraft are , then the labels of the target behavior rules can be set to 211, 213, ..., 21m in sequence.

[0011] Preferably, the specific method for visualizing the target law in step 7 is: Select each cluster The track with the largest number of neighbors is regarded as the characteristic track of the behavior law. Only the characteristic track is drawn in the visualization display, and the target behavior law label corresponding to the characteristic track is marked. The target heading is indicated by an arrow, and the length of the interval between adjacent track points is used to indicate the speed of the target.

[0012] The beneficial effects of the present invention are as follows: a method for mining target behavior patterns based on deep graph clustering provided by the present invention makes full use of the high-dimensional information of the target and the relationship information between targets, and can effectively solve the problems that the mining of target activity patterns based on traditional clustering methods relies on the construction of similarity measurement, and has strong data dependence, high computational complexity, inaccurate measurement description, etc. when facing high-dimensional target data in complex scenes, thereby realizing efficient mining and analysis of target behavior patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is an overall flow chart of the target behavior law mining method based on deep graph clustering of the present invention. DETAILED DESCRIPTION

[0014] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0015] Step 1, set the target's attributes and type labels. This method forms a target behavior regularity label by setting the target attribute, type and cluster number label, thereby representing the regular behavior of the target. In the field of early warning and monitoring, the target attributes are divided into our side, enemy side and friendly side, represented by labels 1, 2 and 3 respectively, and the target types are divided into military aircraft, civilian aircraft, warships and civilian ships, etc., represented by labels 1, 2, 3 and 4 respectively.

[0016] Step 2: Represent the target-related data in the form of a spatiotemporal graph ,in, is a collection of nodes, and Respectively The edge set and adjacency matrix at each moment, the space-time graph is a directed weighted heterogeneous dynamic graph.

[0017] Step 3: Design a spatiotemporal graph autoencoder with an attention mechanism to learn node representations by aggregating adjacency matrix information and reconstruct 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 formula: ; in, , They are the nodes before and after aggregating neighbor information. The expression, Representative Node The neighbor set of Represents the node pair The attention weight between .

[0018] With the node representation, the original network structure can be reconstructed by calculating the inner product of node pairs, thereby achieving unsupervised node representation learning.

[0019] ;

[0020] in, It can be understood as a node pair The probability that there is an edge between them.

[0021] Step 4: Build a self-training graph neural network model, aggregate the nearest neighbor target information after dimensionality reduction, and mine the structural information between targets. Set the cluster center to , then the probability that a node belongs to a certain category , as shown below: ; here, It can be seen as the distribution of node allocation. is the number of nodes. In order to introduce clustering information to realize cluster-oriented node representation, each node is closer to the corresponding cluster center, and the distance within the class is minimized and the distance between classes is maximized. Therefore, the following target distribution is defined: ; Finally, by calculating the difference between the two distributions KL Divergence is used to achieve mutual constraints, which is the so-called self-training.

[0022] ;

[0023] is the loss function of the model.

[0024] node The cluster in which can be calculated by the following formula: ; Step 5: Design a dual self-supervisory module of graph convolutional neural network and deep neural network, build an unsupervised learning mechanism of the model through a two-way mutual supervision mode, and obtain the cluster labels of the node set with the same target.

[0025] Will Considered as labels to supervise the graph convolutional neural network module: ; distributed It can play a bridging role, constraining the representations learned by the deep neural network and the representations learned by the graph convolutional neural network. The complete loss function of the model is as follows: ; , is the weight coefficient.

[0026] node The calculation formula for cluster labels is updated as: ; Step 6: Set the target behavior law label. In the field of early warning and surveillance, the clusters obtained after clustering the multi-dimensional track data of enemy military aircraft are , then the labels of the target behavior rules can be set to 211, 213, ..., 21m in sequence.

[0027] Step 7: Visualize the target behavior. To intuitively represent the target behavior, we can select the track with the largest number of neighbors in each cluster as the characteristic track of the behavior. When visualizing, only the characteristic track is drawn, and the target behavior label corresponding to the characteristic track is marked. The target heading is indicated by an arrow, and the length of the interval between adjacent track points is used to indicate the speed of the target.

[0028] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A target behavior law mining method based on deep graph clustering, characterized in that: The following steps are involved: Step 1, set the target's attributes and type tags; Step 2: Represent the target data in the form of a spatiotemporal graph ; Step 3: Design a spatiotemporal graph autoencoder with an attention mechanism to 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: Build a self-training graph neural network model, aggregate the nearest neighbor target information after dimensionality reduction, and mine the structural information between targets; Step 5: Design a dual self-supervisory module of graph convolutional neural network and deep neural network, build an unsupervised learning mechanism of the model through a two-way mutual supervision mode, and obtain the cluster labels of the target node set; Step 6: Set the target behavior pattern label; Step 7: Visualize the target behavior patterns.

2. The target behavior law mining method based on deep graph clustering according to claim 1 is characterized in that: The spatiotemporal graph autoencoder with attention mechanism described in step 3 is designed to update the aggregated neighbor information model of node representation as shown in the following formula: ； in, , They are the nodes before and after aggregating neighbor information. The expression, Representative Node The neighbor set of Represents the node pair The attention weight between After obtaining the node representation, the original network structure is reconstructed by calculating the inner product of the node pairs, thereby realizing unsupervised node representation learning: ； in, is the probability that an edge exists between a pair of nodes.

3. The target behavior law mining method based on deep graph clustering according to claim 1 is characterized in that: When constructing the self-training graph neural network model in step 4, the cluster center is set to , then the node The probability of belonging to a certain category , as shown below: ； in, As the distribution of node allocation, is the number of nodes. In order to introduce clustering information to realize cluster-oriented node representation, make each node closer to the corresponding cluster center, achieve the minimum intra-class distance and the maximum inter-class distance, define the following target distribution: ； Finally, by calculating the difference between the two distributions KL Divergence is used to achieve mutual constraints, i.e. self-training: ； is the loss function of the model; node The cluster in which can be calculated by the following formula: 。 4. The target behavior law mining method based on deep graph clustering according to claim 1 is characterized in that: When designing the dual self-supervisory module of graph convolutional neural network and deep neural network as described in step 5, Considered as labels to supervise the graph convolutional neural network module: ； distributed It can play a bridging role, constraining the representations learned by the deep neural network and the representations learned by the graph convolutional neural network. The complete loss function of the model is as follows: ； , is the weight coefficient; node The calculation formula for cluster labels is updated as: 。 5. The target behavior law mining method based on deep graph clustering according to claim 1 is characterized in that: The specific method for visualizing the target law described in step 7 is: The track with the largest number of neighbors in each cluster is selected as the characteristic track of the behavior pattern. Only the characteristic track is drawn in the visualization, and the target behavior pattern label corresponding to the characteristic track is marked. The target heading is indicated by an arrow, and the length of the interval between adjacent track points is used to indicate the speed of the target.

Images