A target clustering method and device based on confidence decision tree clustering

Abstract

The present invention proposes a target grouping method and device based on a confidence decision tree, which obtains the individual attributes, capability attributes and communication relationships of the target through detection; takes all targets as root nodes; performs node division from the root node; traverses all nodes to be divided until all nodes do not need to be divided, and finally forms a confidence decision tree; and determines the group to which the target belongs according to the confidence that the target belongs to a certain child node in the confidence decision tree. The present invention determines the cutting point when dividing the nodes of the confidence decision tree during the division process, and proposes a confidence division method based on the cutting point, thereby generating child nodes to finally form a confidence decision tree. Finally, the adversarial group obtained by the confidence decision tree can be adjusted when the number of groups is known. The adversarial group obtained by the present invention can describe the adversarial group through the path of the decision tree, and the confidence division can realize the accurate judgment of the relationship between the adversarial target and the adversarial group.

Description

Technical Field

[0001] The present invention belongs to the technical field of situation analysis, and in particular relates to a target grouping method and device based on confidence decision tree clustering. Background Art

[0002] Target grouping is an important part of situation analysis, providing an information basis for commanders to judge the intention of confrontation. The target grouping task is an understanding and judgment of the organizational structure of the confrontation system by the defender. Its purpose is to find out the relationship between the various groups in the confrontation system, and on this basis, combine their confrontation capabilities to infer their confrontation tasks. In this way, the tactical intentions of the opponent in the entire confrontation environment can be understood. This process can be abstracted as a typical clustering problem.

[0003] The clustering results obtained by existing clustering models, such as density-based methods and entropy-based methods, often lack an intuitive description of the adversarial group, making it difficult for commanders to understand the clustering results when applied to target grouping. In addition, in existing clustering methods, a target can only belong to one adversarial group, so it is difficult to accurately describe some adversarial targets that serve multiple target groups at the same time in the application of target grouping. These adversarial targets often have a "hub" identity in the adversarial system and are important members of the adversarial system.

[0004] Existing clustering-based target grouping methods are difficult to generate adversarial groups that are easy for humans to understand, and it is difficult to accurately divide some targets with complex functionality and participating in multiple adversarial group tasks. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a target grouping method and device based on confidence decision tree clustering. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0006] The present invention provides a target grouping method based on confidence decision tree clustering, comprising:

[0007] Step 1: Obtain the target's individual attributes, capability attributes and communication relationships through detection;

[0008] Step 2, place all targets at the root node;

[0009] Step 3: start dividing the child nodes from the root node. The i-th division process is:

[0010] (1) For a single cluster node, select any attribute of the node and select a possible cutting point in the attribute. The true cutting point is determined by calculating the average confidence contour measure of all possible cutting points and finding its maximum value;

[0011] (2) For the above single cluster node, assume that the node needs to be split to generate three child nodes. The maximum evaluation value is obtained by calculating the confidence contour measurement of all targets at all possible cutting points, and then determine whether the obtained maximum evaluation value is greater than the evaluation value before the division; if so, continue to execute (3); otherwise, there is no need to divide the node;

[0012] If the node is a root node, there is no need to make a judgment and (3) can be directly executed. If so, (3) is continued. Otherwise, there is no need to divide the node.

[0013] (3) generating sub-nodes for the single cluster node partition, and determining the center point of each sub-node according to the real cutting point;

[0014] (4) calculating the confidence of each child node by the distance from all targets to each central node; wherein the confidence expresses the degree to which the target belongs to the group cluster represented by the child node;

[0015] Step 4, traverse all nodes until all nodes do not need to be divided, and finally form a confidence decision tree;

[0016] The traversal order is from top to bottom, and the sibling nodes of the same level are traversed first. If there is no sibling node that can be divided, the lower level nodes are traversed;

[0017] Step 5: According to the confidence that the target belongs to a certain child node in the confidence decision tree, the group cluster to which the target belongs is determined.

[0018] The present invention provides a target grouping device based on confidence decision tree clustering, comprising:

[0019] The detector is used to obtain the target attributes and related functional attributes of the target through detection;

[0020] The executor is used to place all targets at the root node; the child nodes are divided starting from the root node. The i-th division process is:

[0021] (1) For a single cluster node, select any attribute of the node and select a possible cutting point in the attribute. The true cutting point is determined by calculating the average confidence contour measure of all possible cutting points and finding its maximum value;

[0022] (2) For the above single cluster node, assuming that the node needs to be split, the maximum evaluation value is obtained by calculating the confidence contour measurement of all targets at all possible cutting points, and then judging whether the obtained maximum evaluation value is greater than the evaluation value before the split; if so, continue to execute (3); otherwise, there is no need to split the node;

[0023] If the node is a root node, there is no need to make a judgment and (3) can be directly executed. If so, (3) is continued. Otherwise, there is no need to divide the node.

[0024] (3) generating sub-nodes for the single cluster node partition, and determining the center point of each sub-node according to the real cutting point;

[0025] (4) calculating the confidence of each child node by the distance from all targets to each central node; wherein the confidence expresses the degree to which the target belongs to the group cluster represented by the child node;

[0026] Step 4, traverse all nodes until all nodes do not need to be divided, and finally form a confidence decision tree;

[0027] The traversal order is from top to bottom, and the sibling nodes of the same level are traversed first. If there is no sibling node that can be divided, the lower level nodes are traversed;

[0028] The classifier is used to determine the group cluster to which the target belongs according to the confidence that the target belongs to a certain child node in the confidence decision tree.

[0029] The present invention proposes a target grouping method and device based on a confidence decision tree, which obtains the individual attributes, capability attributes and communication relationships of the target through detection; takes all targets as root nodes; performs node division from the root node; traverses all nodes to be divided until all nodes do not need to be divided, and finally forms a confidence decision tree; and determines the group to which the target belongs according to the confidence that the target belongs to a certain child node in the confidence decision tree. The present invention determines the cutting point when dividing the nodes of the confidence decision tree during the division process, and proposes a confidence division method based on the cutting point, thereby generating child nodes to finally form a confidence decision tree. Finally, the adversarial group obtained by the confidence decision tree can be adjusted when the number of groups is known. The adversarial group obtained by the present invention can describe the adversarial group through the path of the decision tree, and the confidence division can realize the accurate judgment of the relationship between the adversarial target and the adversarial group.

[0030] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The present invention provides a flow chart of a target grouping method based on confidence decision tree clustering;

[0032] Figure 2 It is a plane display diagram of the confrontation target provided by the present invention;

[0033] Figure 3 It is a tree structure diagram when performing target grouping tasks. DETAILED DESCRIPTION

[0034] The present invention is further described in detail below with reference to specific embodiments, but the embodiments of the present invention are not limited thereto.

[0035] like Figure 1 As shown, the present invention provides a target grouping method based on confidence decision tree clustering, including:

[0036] Step 1: Obtain the target's individual attributes, capability attributes and communication relationships through detection;

[0037] Step 2, place all targets at the root node;

[0038] Step 3: start dividing the child nodes from the root node. The i-th division process is:

[0039] (1) For a single cluster node, select any attribute of the node and select a possible cutting point in the attribute. The true cutting point is determined by calculating the average confidence contour measure of all possible cutting points and finding its maximum value;

[0040] It is worth noting that: if the communication relationship between targets can be obtained, the average contour metric needs to be replaced by the evaluation index in Equation 15;

[0041] (2) For the above single cluster node, assume that the node needs to be split, calculate the confidence contour measure of all targets at all possible cutting points to obtain the maximum evaluation value, and determine whether the obtained maximum evaluation value is greater than the evaluation value before the split; if so, continue to execute (3), otherwise, there is no need to split the node;

[0042] Among them, in the first method, the maximum evaluation value is the maximum average of the confidence contour metrics of the three child nodes, and the evaluation value before division is the maximum average of the confidence contour metrics of the node at the real cutting point; in the second method, the maximum evaluation value is calculated based on the maximum average of the confidence contour metrics of the three child nodes using formula 15; the evaluation value before division is obtained based on the maximum average of the confidence contour metrics of the node at the real cutting point using formula 15. The second method is suitable for the case where the communication relationship can be obtained.

[0043] It is worth noting that: if the node is the root node, no judgment is required and (3) can be executed directly.

[0044] (3) generating sub-nodes for the single cluster node partition, and determining the center point of each sub-node according to the real cutting point;

[0045] (4) calculating the confidence of each child node by the distance from all targets to each central node; wherein the confidence expresses the degree to which the target belongs to the group cluster represented by the child node;

[0046] Step 4, traverse all single cluster leaf nodes until all nodes do not need to be divided, and finally form a confidence decision tree;

[0047] The traversal order is from top to bottom, with sibling nodes of the same level being traversed first. If there are no sibling nodes that can be divided, the lower-level nodes are traversed.

[0048] Step 5: According to the confidence that the target belongs to a certain child node in the confidence decision tree, the group cluster to which the target belongs is determined.

[0049] The present invention proposes a target grouping method and device based on a confidence decision tree, which obtains the individual attributes, capability attributes and communication relationships of the target through detection; takes all targets as root nodes; performs node division from the root node; traverses all nodes to be divided until all nodes do not need to be divided, and finally forms a confidence decision tree; and determines the group to which the target belongs according to the confidence that the target belongs to a certain child node in the confidence decision tree. The present invention determines the cutting point when dividing the nodes of the confidence decision tree during the division process, and proposes a confidence division method based on the cutting point, thereby generating child nodes to finally form a confidence decision tree. Finally, the adversarial group obtained by the confidence decision tree can be adjusted when the number of groups is known. The adversarial group obtained by the present invention can describe the adversarial group through the path of the decision tree, and the confidence division can realize the accurate judgment of the relationship between the adversarial target and the adversarial group.

[0050] In a specific implementation of the present invention, in the i-th division process (1) includes:

[0051] (11) For a single cluster node, select any attribute of the node;

[0052] (12) Traversing all possible cutting points in the selected target attributes, and calculating the confidence contour metric of all targets after the attributes are cut according to each possible cutting point;

[0053] (13) Calculate the average value of the confidence contour metric at each cutting point, and select the cutting point with the largest average value of the confidence contour metric as the true cutting point.

[0054] To construct a decision tree, it is necessary to find the best partitioning attribute and the associated cut point at each node so that the splits are made continuously, where the tree is initialized with a root node containing the attribute values ​​of all targets. The key question is how to evaluate and find the best cut point for the split. In the field of unsupervised learning, the silhouette metric is a popular internal criterion for evaluating the quality of the partition, which takes into account both cohesion (the average distance between the target and all other targets in the same adversarial group) and separation (the distance between the target and all other targets in the nearest adversarial group). The silhouette metric was originally defined for hard partitioning, which is not applicable to the confident partitioning of nodes. Therefore, it is necessary to extend the original silhouette metric in the framework of confident partitioning.

[0055] Definition 1 (Confidence Contour Metric) For any target x i The confidence silhouette measure (ES) of ∈X is defined as

[0056]

[0057] Where a(x i ) represents confidence cohesion, calculated as x i With the same opposing groupω s The weighted average distance between all other targets within:

[0058]

[0059] And b(x i ) represents confidence separation, calculated as x i With the recent adversarial groupω n The weighted average distance between all other targets in :

[0060]

[0061] Among them BetP j (ω s ) and BetP j (ω n ) represents the target x j Belong to the antagonistic group ω s and adversarial groups n The pignistic probability of proposition A can be obtained by the following formula:

[0062]

[0063] where || represents the cardinality of the set V. The confidence contour metric defined above covers the original contour metric defined for hard segmentation. j (ω s ) and BetP j (ωn ) takes 1 (i.e., the adversarial group confidence is certain), it will degenerate into the original profile metric. In addition, similar to the original profile metric, the confidence profile metric defined above also takes values ​​between -1 and 1. i When the contour metric value of is close to 1, it contains x i The adversarial group is compact, and x i Stay away from other adversarial groups, which is a good situation. However, when x i The contour metric value is negative (i.e. b(x i ) <a(x i )) This means that x i Compared with the target in another adversarial group, x i Targets of the same adversarial group are closer. In many cases, this is an unfavorable situation and should be avoided. In order to evaluate the quality of the segmentation using each cut point, the average confidence profile (AES) value of all targets in the dataset needs to be calculated as follows

[0064]

[0065] Among them BetP j (ω s ) is the target x i Belong to its opposing groupω s The pignistic probability of the adversarial group is the largest pignistic probability. For each node that needs to be divided, the average confidence contour metric value of all possible cutting points is calculated by traversing all attributes. The present invention will select the cutting point with the largest average confidence contour metric value, and the cutting point will be used for confidence partitioning.

[0066] In a specific implementation of the present invention, in the i-th division process (2) includes:

[0067] (21) For the single cluster node, assuming that the node needs to be split, calculate the confidence contour metrics of all targets at all possible cutting points;

[0068] (22) For the single cluster node, calculate the average value of the confidence contour measure after splitting at each possible cutting point;

[0069] (23) Select the maximum average value from the average values ​​of the confidence contour metrics obtained in (22), obtain the maximum evaluation value based on the maximum average value, and determine whether the obtained maximum evaluation value is greater than the evaluation value before the division. If so, the node can be split. Otherwise, there is no need to split the node.

[0070] Using the cutting point selected in (2), the present invention needs to test in advance whether the contour metric value after the division is greater than the previous value. If so, the node to be divided will be divided using the selected cutting point to generate a new child node; otherwise, there is no need to split this node.

[0071] In the developed unsupervised confidence decision tree, there are two types of nodes: single cluster nodes and composite cluster nodes (i.e., the union of single clusters). This implementation will describe how to perform confidence division on these two types of nodes.

[0072] In a specific implementation of the present invention, in the i-th division process (3) includes:

[0073] (31) If the condition in (2) is met, the generation of child nodes is executed;

[0074] (32) If the child node is a single cluster node, the center point of the single cluster node is calculated according to the single cluster node center calculation expression;

[0075] (33) If the child node is a composite cluster node, the true cutting point is determined as the center point of the composite cluster node, and the center points of the remaining single cluster nodes are calculated.

[0076] In a specific implementation of the present invention, in the i-th division process (4) includes:

[0077] (41) Calculate the distance from all targets to the center point corresponding to each child node;

[0078] (42) Based on the distance in (41), calculate the confidence of each child node.

[0079] In a specific embodiment of the present invention, in the i-th division process before (3), the target grouping method based on confidence decision tree clustering further includes:

[0080] a1, determine whether the node is a single cluster node or a composite cluster node;

[0081] b1, if the node is a single cluster node, the actual cutting point in (1) is used as the cutting point for generating child nodes of the node;

[0082] 1) Confidence partitioning of single cluster nodes

[0083] Node A represents a single adversarial group to be split, and the child nodes after the split are denoted as L, R, and L∪R. At the beginning, the centers of these child nodes need to be estimated. Intuitively, the center of the composite node L∪R is m can be represented by the selected cut point at node A. Therefore, the centers of the left child node L and the right child node R are calculated as

[0084]

[0085]

[0086] Where Left = {x i |x i ∈A,x ij <Center m} and Righ={x i |x i ∈A,x ij >Center m} represents the left set and the right set separated by the cut point, where x ij is the target on the j-th attribute selected for splitting. i (A) is the target x i The confidence membership of node A.

[0087] In order to eliminate the influence of different adversarial group shapes on the mass function results, the normalized distance is used as

[0088]

[0089]

[0090]

[0091] Where γ>0 is a parameter that adjusts the importance of composite clusters. Larger values ​​of γ will penalize the membership of composite clusters, and when γ approaches infinity, the above confidence membership will only be reduced to fuzzy membership. The default value of 1 is recommended to obtain appropriate composite clusters for characterizing overlaps between adversarial groups. The mass function m constructed in Eq. 6 is i , A confidence partition of the target in node A to its three child nodes is provided, where the composite cluster represented by the middle child node characterizes the overlap between the left and right individual adversarial groups. In addition, based on these mass functions, the boundaries between these generated adversarial groups can also be determined to make the confidence partition results easier for users to interpret.

[0092] Then, for each target x in node A i The confidence membership to the three child nodes is inversely proportional to the distance to the center of the corresponding adversarial group. Therefore, i The mass function of the confidence membership can be designed as:

[0093]

[0094]

[0095]

[0096] in, They are x on the selected elements. i To Center l 、Center r and Center m distance.

[0097] c1, if the node is a composite cluster node, its two sibling nodes need to be divided;

[0098] d1, obtain at least one real cutting point of a sibling node in the process of dividing the child node;

[0099] Among them, if both brother nodes cannot be divided, then the node does not need to be divided;

[0100] e1, use the true cut obtained in d as the true cut point of the composite cluster node to generate the child node, and calculate the center and confidence of its child node at one time.

[0101] 2) Confidence Partitioning of Composite Cluster Nodes

[0102] For the confidence partition of the node representing the composite cluster, there is no need to find a new cutting point at this node. Instead, the present invention will use the boundaries calculated from its sibling nodes to directly form the confidence partition. Therefore, before this step, it is necessary to ensure that the splitting of its sibling nodes has been completed. Since the composite cluster is the union of the adversarial groups represented by its sibling nodes, the boundaries of the sibling nodes can be inherited to split the composite cluster. If multiple sibling nodes generate new child nodes, the boundaries of these sibling nodes will be used one by one.

[0103] In a specific implementation of the present invention, if only one of the sibling nodes in d needs to generate a child node, then the number of real cutting points obtained is 1; if both sibling nodes need to generate child nodes, then the number of real cutting points obtained is 2;

[0104] When there are two real cutting points in the process of partitioning the composite cluster nodes, any one of the real cutting points can be selected to perform a partition to generate two child nodes, and another real cutting point can be selected to perform a partition on each of the two generated child nodes to generate two child nodes.

[0105] For nodes with composite clusters, using the above partitioning method, the number of leaf nodes generated by the entire decision tree is exponential in the number of different adversarial groups. For a dataset consisting of K different adversarial groups, the tree will generate at most 2 Kleaf nodes, however, most of the leaf nodes representing composite clusters composed of many single adversarial groups are meaningless, which also complicates the structure of the tree. To solve this problem, the composite clusters can be directly partitioned using the cutting points obtained from their sibling nodes, which can prevent the generation of composite clusters composed of more than two single adversarial groups. With this simplified confidence partitioning rule, the number of generated leaf nodes is reduced to the quadratic size of the number of different adversarial groups. This simplified decision tree is preferred in applications due to its low complexity and better interpretability without sacrificing too much flexibility in confidence partitioning.

[0106] In a specific embodiment of the present invention, after step 4, the target grouping method based on confidence decision tree clustering further includes:

[0107] a2, determining in the confidence decision tree that there is no need to divide a child node, and treating the child node as a leaf node;

[0108] b2, determine whether the number of leaf nodes is greater than or less than a given number;

[0109] b3: If it is greater than or less than a given number, the leaf nodes in the confidence decision tree are adjusted until it is consistent with the given number.

[0110] The procedures developed in the above two parts are initially executed at the root node and will be repeated until all leaf nodes have been evaluated and are not suitable for splitting. At this point, a preliminary unsupervised confidence decision tree has been constructed, and the leaf nodes (nodes of separate clusters) represent the generated adversarial groups. If the number of adversarial groups is unknown, or it happens to be equal to the number obtained from the generated tree, the final clustering result can be obtained directly from the tree. Otherwise, the tree needs to be adjusted for adversarial groups to adapt to the given number of adversarial groups.

[0111] In a specific embodiment of the present invention, b3 comprises:

[0112] If the number of leaf nodes is greater than the given number, the leaf node with the worst confidence is merged into the leaf node with the closest distance to obtain a merged node;

[0113] Determine the composite cluster node of the two leaf nodes to be merged;

[0114] Merge the merged node with the composite cluster node to reduce the number of single clusters by one, and repeat the above process until the number of leaf nodes is consistent with the given number;

[0115] If the number of leaf nodes is less than the given number, try to split the current leaf node. During the splitting process, there is no restriction that the maximum evaluation value must be greater than the current evaluation value.

[0116] The partitioning method with the largest evaluation value is selected until the number of leaf nodes reaches the given number.

[0117] There are two possible situations where adversarial group adjustment is required. First, the number of adversarial groups is greater than the given adversarial group. In this case, the generated adversarial groups need to be merged. Each generated single adversarial group ω j The average confidence profile (AES) value of all targets in is used to measure the quality of the adversarial group.

[0118]

[0119] Among them, BetP i (ω j ) is the target x i Belong to its opposing groupω j The probability of ES(x i ) is the target x defined in Eq. 1 i The confidence profile metric of . Using the above evaluation metric, the single adversarial group with the worst quality is merged with its closest single adversarial group (measured by the Euclidean distance between the centers of the two adversarial groups). At the same time, those composite clusters that contain the merged adversarial groups will also be merged together. For example, if adversarial group A is merged with adversarial group B, then composite cluster A∪C should also be merged with B∪C. This step will continue until the specified number of adversarial groups is reached.

[0120] The second case is when the number of adversarial groups is less than the given adversarial group. In this case, further partitioning is required. All possible partitioning methods are tested on the leaf nodes with a single adversarial group, and the partitioning method with the maximum average confidence silhouette metric value among all targets in the dataset is selected. This step will continue until the specified number of adversarial groups is reached.

[0121] The present invention provides a target grouping device based on confidence decision tree clustering, comprising:

[0122] The detector is used to obtain the target attributes and related functional attributes of the target through detection; the executor is used to take all targets as root nodes; the sub-nodes are divided from the root node, and the i-th division process is:

[0123] (1) For a single cluster node, select any attribute of the node and select a possible cutting point in the attribute. The true cutting point is determined by calculating the average confidence contour measure of all possible cutting points and finding its maximum value;

[0124] Among them, if the communication relationship between targets can be obtained, the average contour metric needs to be replaced by the evaluation index in formula 15;

[0125] (2) For the above single cluster node, assume that the node needs to be split to generate three child nodes. The maximum evaluation value is obtained by calculating the confidence contour measurement of all targets at all possible cutting points, and then determine whether the obtained maximum evaluation value is greater than the evaluation value before the division; if so, continue to execute (3); otherwise, there is no need to divide the node;

[0126] If the node is a root node, there is no need to make a judgment and (3) can be directly executed. If so, (3) is continued. Otherwise, there is no need to divide the node.

[0127] (3) If the conditions in (2) are met, the generation of child nodes is executed, and the center point of each child node is determined according to the actual cutting point;

[0128] (4) calculating the confidence of each child node by the distance from all targets to each central node; wherein the confidence expresses the degree to which the target belongs to the group cluster represented by the child node;

[0129] Traverse all single-cluster leaf nodes until all nodes do not need to be divided, and finally form a confidence decision tree;

[0130] The traversal order is from top to bottom, and the sibling nodes of the same level are traversed first. If there are no sibling nodes that can be divided, the lower-level nodes are traversed.

[0131] The classifier is used to determine the group cluster to which the target belongs according to the confidence that the target belongs to a certain child node in the confidence decision tree.

[0132] The effect of the present invention was evaluated and verified on the problem of target grouping in a confrontation environment.

[0133] (1) Problem description

[0134] The main task of target clustering is to divide the adversarial targets into different adversarial groups based on the adversarial target attributes detected by sensors and the communication links between adversarial targets. The adversarial target attributes detected by sensors will be used as attributes for the confidence decision tree to be divided at the nodes, such as the three-dimensional coordinates of the target, the angle with the x-axis, the angle with the y-axis, the speed, the air defense capability, the anti-ship capability, the ground attack capability, the reconnaissance and early warning, the electromagnetic suppression, the communication capability, the obstacle clearance capability, etc. The adversarial capability-related attributes of the target can be digitized according to the strength of the capability to facilitate the selection of cutting points when the decision tree is divided. The communication links between the intercepted adversarial targets can be used as one of the objective functions to penalize the division methods that have communication links but are not classified into the same cluster. The present invention adjusts the objective function for calculating the evaluation value of a single target in Formula 1 to obtain the following formula:

[0135]

[0136] Among them, Node ＝ Is the node where X is located, Node ≠ is the target group closest to X, Re(') is the set of targets associated with target X, θ(X') is the relationship between X' and Node ≠ γ is the proportion of the target profile metric to its evaluation value, which can be adjusted according to the importance of the association relationship in the evaluation. The value of γ is between [0,1]. The smaller the value, the greater the proportion of the association relationship in the evaluation. The part of evaluating the target through the association relationship is composed of the second part of formula 9, where the fraction represents the ratio of the membership of the target that has an association relationship with the target and is in the same cluster to the membership of all the targets that have an association relationship. If the communication relationship cannot be obtained, it can be calculated according to formula 1.

[0137] (2) Experimental results

[0138] The tree structure obtained by the confidence decision tree is as follows Figure 3 As shown (since some leaf nodes do not have the target with the highest confidence for the node, such nodes are not shown in the figure).

[0139] The present invention generates six confrontation groups and the union of any two of them by performing confidence division on the anti-submarine capability (F3), speed (F14), X-axis (F9), Y-axis (F10), and Z-axis (F11) of the confrontation target, thereby obtaining the division results and their corresponding confidence levels. The final grouping results of the confrontation target in the plane are distinguished by color. Figure 2 The adversarial targets included in each adversarial group belong to the adversarial group with the highest confidence. The specific adversarial targets and their confidences included in each adversarial group are shown in Table 1.

[0140] Table 1 Target intelligence information

[0141]

[0142]

[0143] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0144] Although the present application is described herein in conjunction with various embodiments, in the process of implementing the claimed application, those skilled in the art may understand and implement other variations of the disclosed embodiments by reviewing the drawings, the disclosure, and the appended claims. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality of components or steps.

[0145] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A target clustering method based on confidence decision tree clustering, characterized in that: include: Step 1: Obtain the target's individual attributes, capability attributes and communication relationships through detection; Step 2, place all targets at the root node; Step 3: start dividing the child nodes from the root node. The i-th division process is: (1) For a single cluster node, select any attribute of the node and select a possible cutting point in the attribute. The true cutting point is determined by calculating the average confidence contour measure of all possible cutting points and finding its maximum value; (2) For the above single cluster node, assume that the node needs to be split to generate three child nodes. The maximum evaluation value is obtained by calculating the confidence contour measurement of all targets at all possible cutting points, and then determine whether the obtained maximum evaluation value is greater than the evaluation value before the division; if so, continue to execute (3); otherwise, there is no need to divide the node; If the node is a root node, there is no need to make a judgment and (3) can be directly executed. If so, (3) is continued. Otherwise, there is no need to divide the node. (3) generating sub-nodes for the single cluster node partition, and determining the center point of each sub-node according to the real cutting point; (4) calculating the confidence of each child node by the distance from all targets to each central node; wherein the confidence expresses the degree to which the target belongs to the group cluster represented by the child node; Step 4, traverse all nodes until all nodes do not need to be divided, and finally form a confidence decision tree; The traversal order is from top to bottom, and the sibling nodes of the same level are traversed first. If there is no sibling node that can be divided, the lower level nodes are traversed; Step 5: According to the confidence that the target belongs to a certain child node in the confidence decision tree, the group cluster to which the target belongs is determined.

2. The target clustering method based on confidence decision tree clustering according to claim 1, characterized in that: In the i-th partitioning process (1) includes: (11) For a single cluster node, select any target attribute of the node; (12) Traversing all possible cutting points in the selected target attributes, and calculating the confidence contour metric of all targets after the attributes are cut according to each possible cutting point; (13) Calculate the average value of the confidence contour metric at each cutting point, and select the cutting point with the largest average value of the confidence contour metric as the true cutting point.

3. The target clustering method based on confidence decision tree clustering according to claim 1, characterized in that: In the i-th partitioning process (2), it includes: (21) For the single cluster node, assuming that the node needs to be split, calculate the confidence contour metrics of all targets at all possible cutting points; (22) For the single cluster node, calculate the average value of the confidence contour measure after splitting at each possible cutting point; (23) Select the maximum average value from the average values ​​of the confidence contour metrics obtained in (22), obtain the maximum evaluation value based on the maximum average value, and determine whether the obtained maximum evaluation value is greater than the evaluation value before the division. If so, the node can be split. Otherwise, there is no need to split the node.

4. The target clustering method based on confidence decision tree clustering according to claim 1, characterized in that: In the i-th partitioning process (3), it includes: (31) If the condition in (2) is met, the generation of child nodes is executed; (32) If the child node is a single cluster node, the center point of the single cluster node is calculated according to the single cluster node center calculation expression; (33) If the child node is a composite cluster node, the true cutting point is determined as the center point of the composite cluster node, and the center points of the remaining single cluster nodes are calculated.

5. The target clustering method based on confidence decision tree clustering according to claim 1, characterized in that: In the i-th partitioning process (4), it includes: (41) Calculate the distance from all targets to the center point corresponding to each child node; (42) Based on the distance in (41), calculate the confidence of each child node.

6. The target clustering method based on confidence decision tree clustering according to claim 1, characterized in that: In the i-th division process before (3), the target clustering method based on confidence decision tree clustering further includes: a1, determine whether the node is a single cluster node or a composite cluster node; b1, if the node is a single cluster node, the actual cutting point in (1) is used as the cutting point for generating child nodes of the node; c1, if the node is a composite cluster node, its two sibling nodes need to be divided; d1, obtain at least one real cutting point of a sibling node in the process of dividing the child node; Among them, if both brother nodes cannot be divided, then the node does not need to be divided; e1, uses the true cut obtained in d1 as the true cut point of the composite cluster node to generate the child node, and calculates the center and confidence of its child node at one time.

7. The target clustering method based on confidence decision tree clustering according to claim 6, characterized in that: In d, if only one of the sibling nodes needs to generate a child node, the actual cutting point obtained is 1; if both sibling nodes need to generate child nodes, the actual cutting points obtained are 2; When there are two real cutting points in the process of partitioning the composite cluster nodes, any one of the real cutting points can be selected to perform a partition to generate two child nodes, and another real cutting point can be selected to perform a partition on each of the two generated child nodes to generate two child nodes.

8. The target clustering method based on confidence decision tree clustering according to claim 1, characterized in that: After step 4, the target grouping method based on confidence decision tree clustering further includes: a2, determining a child node that does not need to be divided in the confidence decision tree, and making the child node a leaf node; b2, determine whether the number of leaf nodes is greater than or less than a given number; b3: If it is greater than or less than a given number, the leaf nodes in the confidence decision tree are adjusted until it is consistent with the given number.

9. The target clustering method based on confidence decision tree clustering according to claim 8, characterized in that b3 include: If the number of leaf nodes is greater than the given number, the leaf node with the worst confidence is merged into the leaf node with the closest distance to obtain a merged node; Determine the composite cluster node of the two leaf nodes to be merged; Merge the merged node with the composite cluster node to reduce the number of single clusters by one, and repeat the above process until the number of leaf nodes is consistent with the given number; If the number of leaf nodes is less than the given number, try to split the current leaf node. During the splitting process, there is no restriction that its maximum evaluation value must be greater than the evaluation value before the split. The partitioning method with the largest evaluation value is selected until the number of leaf nodes reaches the given number.

10. A target clustering device based on confidence decision tree clustering, characterized in that: include: The detector is used to obtain the target attributes and related functional attributes of the target through detection; Executor, which is used to place all targets at the root node; Starting from the root node, the child nodes are divided. The i-th division process is: (1) For a single cluster node, select any attribute of the node and select a possible cutting point in the attribute. The true cutting point is determined by calculating the average confidence contour measure of all possible cutting points and finding its maximum value; (2) For the above single cluster node, assume that the node needs to be split to generate three child nodes. The maximum evaluation value is obtained by calculating the confidence contour measurement of all targets at all possible cutting points, and then determine whether the obtained maximum evaluation value is greater than the evaluation value before the division; if so, continue to execute (3); otherwise, there is no need to divide the node; If the node is a root node, there is no need to make a judgment and (3) can be directly executed. If so, (3) is continued. Otherwise, there is no need to divide the node. (3) generating sub-nodes for the single cluster node partition, and determining the center point of each sub-node according to the real cutting point; (4) calculating the confidence of each child node by the distance from all targets to each central node; wherein the confidence expresses the degree to which the target belongs to the group cluster represented by the child node; Step 4, traverse all nodes until all nodes do not need to be divided, and finally form a confidence decision tree; The traversal order is from top to bottom, and the sibling nodes of the same level are traversed first. If there is no sibling node that can be divided, the lower level nodes are traversed; The classifier is used to determine the group cluster to which the target belongs according to the confidence that the target belongs to a certain child node in the confidence decision tree.

Images