CN114330509A

CN114330509A - Method for predicting activity rule of aerial target

Info

Publication number: CN114330509A
Application number: CN202111511600.0A
Authority: CN
Inventors: 柳伟; 徐焕祥; 谈民; 程义; 许凯钰; 侯卫华
Original assignee: Zhongke Star Map Co ltd
Current assignee: Zhongke Star Map Co ltd
Priority date: 2021-12-06
Filing date: 2021-12-06
Publication date: 2022-04-12

Abstract

The invention provides a method for predicting an activity rule of an aerial target, which comprises the following steps: step 1, pretreatment stage: screening the original data of investigation/training/patrol, and simultaneously carrying out normalization processing and formatting processing on target track information to form a standard data format required by modeling; step 2, generating a preliminary track: obtaining a segmented track according to an improved DBSCAN method for the track; step 3, extracting the multi-dimensional characteristics of the segmented track: updating the multidimensional characteristics of the track according to the height change, speed change, direction angle change, acceleration change and airspace change of the segmented track calculation target, and storing the multidimensional characteristics in a multidimensional track sample library; step 4, constructing a Catboost classifier: dividing a training set, a verification set and a test set based on the accumulated data, generating a reconnaissance/training/patrol rule model, carrying out reconnaissance/training/patrol track updating, and storing the track into a route model library; and predicting the target activity rule by using the constructed Catboost classifier.

Description

Method for predicting activity rule of aerial target

Technical Field

The invention relates to the technical field of target detection and identification, in particular to an air target activity rule prediction method.

Background

Obtaining decision advantages and obtaining battlefield control power is one of the key conditions for information-based war winning. However, in an informatization combat environment, target objects are numerous, collaborative relationships assist and maneuver frequently. This makes the commander unable to make timely and effective decisions in the face of the vast information and rapidly changing battlefield situations.

In order to assist the commander to implement the combat command and further obtain the decision advantages, the command decision system is required to be utilized to perform information fusion, namely, the multi-source information is quickly and efficiently processed and abstracted so as to perform situation assessment subsequently. By target grouping, the number of targets concerned by the commander can be greatly reduced, and the cognitive pressure faced by the commander is reduced.

And the historical data is deeply mined and learned by adopting methods such as data mining, machine learning and the like, the rules of target group activities are mined and analyzed from massive historical data and are converted into knowledge for storage, and reliable bases can be more effectively provided for relationship identification and situation estimation of task groups.

The situation estimation is mainly described in the definition of functions to be realized by the data fusion technology, including estimation of the position of an enemy in a military application scene, identification of the identity of the enemy, situation assessment and threat estimation after primary fusion processing.

The situation assessment mainly comprises the analysis of the battle activities of the enemy, battlefield events, maneuvering modes, the position of the enemy and the force organization of the enemy, and the situation reasoning and threat estimation are carried out according to the battlefield situation information and the target information, so that the objective assessment of the enemy target is expected. The situation assessment process is mainly based on the current combat environment, combines expert experience knowledge and instant information of target entities, and generates high-level information description aiming at enemies according to the mutual relation between the current task target and the battlefield entity monitored by the sensor. The final purpose of situation assessment is to determine existence of the moving entity and the target regular information thereof and the possibility of the state of the moving entity and the target basic information thereof according to the monitored information of the moving entity and the target regular information in the battlefield environment, so that the cognition of the enemy is improved. And finally, forming a multi-angle and multi-level battlefield situation analysis view through target grouping, carrying out short-term prediction on possible situation development, carrying out certain threat estimation on an enemy combat entity, and guiding fire distribution.

The analysis and mining of the activity rules need to analyze historical data based on a mathematical statistics method, extract the frequently-occurring range of the target and the typical track of the target, and analyze the typical track of the target. Clustering is an important data mining technology, a data set can be divided into a plurality of clusters, objects in the same cluster have high proximity, and objects which cannot be added into any cluster are regarded as isolated points. The original track point data density is high, and if the original track data points are all adopted for track prediction, the prediction efficiency is low due to large calculation amount. The method is characterized in that the original track points in a certain period of time are clustered by adopting a DBSCAN method, and the track points capable of representing the track in the period of time are selected, so that the calculated amount of the track points for predicting the track can be reduced, the prediction efficiency of the track prediction points is improved, and the macroscopic display of the track prediction is facilitated.

The density-based clustering algorithm takes local data characteristics as a clustering judgment standard, and the main idea is as follows: clustering continues as long as the density of the neighboring regions exceeds a certain threshold. A class is considered to be a region of data in which objects are dense and regions in which objects are sparse separate classes. The density-based clustering algorithm does not need to preset the number of clusters, but finds the clusters of any shape and size through the data density, namely the number of objects in a unit area, thereby overcoming the defect that the distance-based algorithm can only find the clusters of 'circular-like'. The research of a density-based clustering algorithm is a very active field, mainly used for clustering spatial data, and DBSCAN is a typical representative thereof.

The DBSCAN algorithm divides areas with a sufficiently high density into clusters, defines the clusters as a maximum set of density connections, controls the growth of the clusters according to a threshold, and is a typical density clustering algorithm based on high density connection areas. DBSCAN represents density in terms of the number of objects present in an epsilon neighborhood, and is defined primarily in several terms as follows.

Defining 1. epsilon neighborhood, given a set of data objects D and a data object p, p ∈ D, epsilon neighborhood N of p_ε(p) is defined as N_ε(p) { q ∈ D | dist (p, q) ≦ epsilon }, where dist (p, q) represents the distance between the two data objects p and q in D.

Definition 2 core objects and boundary objects, for a data object p ∈ D given the integer MinPts, if the number of objects | N of the ε neighborhood of p_ε(p) | is greater than or equal to MinPts, then p is called a core object; if | N_ε(p)|<MinPts, but q is the core object p ∈ N_εAnd (q), then p is called a boundary object.

Definition 3. direct Density is achievable if p, p ∈ D, for a given ε and MinPts, there is p ∈ N_ε(q)，|N_ε(p) | ≧ MinPts, object p is said to be reachable from object q with respect to ε and MinPts direct densities.

Definition 4. Density is achievable if p₁，p₂，…，p_n∈D，p₁＝q,p_nP, p for a given ε and MinPts_i+1From p_iDirect density is reachable, then object p is said to be reachable from object q with respect to ε and MinPts densities.

Definition 5. density is connected, if p, q ∈ D, for a given epsilon and MinPts,

such that object p and object q are both reachable from o density, then object p and object q are said to be connected with respect to ε and MinPts densities。

Definition 6. clusters and outliers, for a given ε and MinPts, a cluster C refers to a non-empty subset of D that satisfies the following condition:

(1)

if p ∈ C and q is reachable from p with respect to ε and MinPts densities, then q ∈ C.

(2)

q and p are connected with respect to ε and MinPts density. Objects that are not contained in any cluster are called outliers.

The DBSCAN algorithm finds clusters and outliers in dataset D according to the above definition, first examines the ε neighborhood of each object in D, and if object p is a core object, generates a table containing N_ε(p) cluster C of objects; then detecting N of all unprocessed objects q in C_ε(q), if q is a core object, N that will not have been included in C_εThe objects in (q) are added to cluster C and the epsilon neighborhood of these objects is detected in the next step. This process is repeated until no new objects are added to the current cluster C.

Although the DBSCAN algorithm has the advantages of mining arbitrary shape clusters and effectively detecting isolated points, the following improvements are still needed in the track prediction application:

(1) the density threshold ε and MinPts are globally unique and difficult to select. If the density threshold value is too large, a cluster is divided into a plurality of clusters, and a large number of isolated points appear; if the density threshold is too small, several clusters that are further apart will be merged.

(2) The algorithm operates on the entire data set, requiring large memory support and I/O consumption when the amount of data is large.

Disclosure of Invention

The method aims to reduce clustering time and memory consumption and establish a target activity rule model. The technical scheme of the invention is as follows: an air target activity rule prediction method comprises the following steps:

step 1, pretreatment stage: screening the original data of investigation/training/patrol, and simultaneously carrying out normalization processing and formatting processing on target track information to form a standard data format required by modeling;

step 2, generating a preliminary track: obtaining a segmented track according to an improved DBSCAN method for the track;

step 3, extracting the multi-dimensional characteristics of the segmented track: updating the multidimensional characteristics of the track according to the height change, speed change, direction angle change, acceleration change and airspace change of the segmented track calculation target, and storing the multidimensional characteristics in a multidimensional track sample library;

step 4, constructing a Catboost classifier: dividing a training set, a verification set and a test set based on the accumulated data, generating a reconnaissance/training/patrol rule model, carrying out reconnaissance/training/patrol track updating, and storing the track into a route model library; and predicting the target activity rule by using the constructed Catboost classifier.

Further, in step 2, the data of the target object in the post-standard data format is grouped according to the designated number of groups, each group is processed respectively to select a core object, a core object list of each group is obtained, then, local clustering is performed according to the core object list, representative points of the local clustering are selected, and finally, global confirmation is performed to obtain the clustering.

Further, the improved DBSCAN algorithm in step 3 includes the following steps:

(1) given a set of data objects D and a data object p, p ∈ D, and defining the epsilon neighborhood N of p_ε(p) is N_ε(p) { q ∈ D | dist (p, q) ≦ epsilon }, where dist (p, q) represents the distance between two data objects p and q in D, and dynamically setting a parameter epsilon by using the non-spatial attribute of the data, making epsilon ═ tv, where t is the set maneuvering time threshold and v is the maneuvering speed;

(2) the data set is scanned and an arbitrary data object p is selected. If p has been classified as a cluster or has been marked as noise, then the data set is scanned again to select a data object; otherwise, judging whether the point in the field is smaller than MinPts, if so, marking the point p as a boundary point or a noise point; otherwise, marking the p point as a core point, establishing a new cluster C and adding all the p point fields into the new cluster C;

(3) aiming at the problem that the core object is qualified for expansion, once a certain core object is found, whether each point in the neighborhood of the certain core object is the core object needs to be judged so as to further expand outwards; selecting a representative object for each region to reduce clustering time and memory consumption;

(4) and checking all the unmarked q points in the neighborhood epsilon, judging whether the points in the neighborhood are smaller than MinPts, if so, continuously checking the unmarked q points in the neighborhood, and otherwise, adding the points which are not classified into other clusters in the neighborhood into C.

Further, specifically, the main purpose of spatial clustering is to reduce the number of spatial target objects with excessive attention, and further optimizing is as follows: the method comprises the steps of firstly carrying out region division grouping on target objects according to static and dynamic data of positions, then carrying out local clustering on each group, and then judging from the whole situation so as to further reduce the memory.

Further, in step 3, the changes of the characteristic quantities such as height, speed, direction angle, acceleration, etc. are measured for the time interval t, the values of these quantities are measured by various sensors at the previous moment, and are measured again at the next moment, and the difference between the two moments of the variables is the change of the characteristic.

Further, the step 4 adopts a CATBOOST method to perform classification training, firstly, representative data obtained by the improved DBSCAN method is divided into a training set and a verification set test set; training parameters by using a training set, and adjusting a hyper-parametric test set by using a verification set to test the performance of the algorithm to obtain a mathematical model for classifying track data; and classifying a data table consisting of position information, type information and activity information in the flight path data by adopting a CATBOOST algorithm, and finally predicting the target activity rule by utilizing a trained model.

Has the advantages that:

(1) the method trains the fast track model by clustering the representative track points, and has the advantages of high calculation speed and high accuracy.

(2) In spatial clustering, the non-spatial attributes of the target object are utilized, that is, a dynamic parameter setting is adopted such that ∈ tv, where t is a set maneuvering time threshold value, and v is a maneuvering speed. Therefore, the situation that when the density threshold epsilon and the MinPts are globally unique and the density threshold is too large, a cluster is divided into a plurality of clusters, and a large number of isolated points appear is avoided; too small a density threshold has the disadvantage of merging several clusters that are far apart.

(3) And selecting the representative object during clustering so as to reduce clustering time and memory consumption. And judgment repetition caused by neighborhood overlapping is avoided.

(4) Grouping in advance during clustering, grouping target objects in advance according to static and dynamic data, respectively processing each group to select core objects, and obtaining a core object list of each group. And then local clustering is performed according to the list. And finally, global judgment is carried out, so that the operation consumption can be further reduced, and the clustering degree is improved.

Drawings

FIG. 1: the invention relates to an activity rule modeling flow chart.

Detailed Description

The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, rather than all embodiments, and all other embodiments obtained by a person skilled in the art based on the embodiments of the present invention belong to the protection scope of the present invention without creative efforts.

Information fusion: also called data fusion, multi-sensor fusion, can be defined as an information processing process using computer technology to automatically analyze and comprehensively infer a plurality of sensor observation information obtained in time sequence under certain criteria so as to complete task decision and evaluation.

And (3) situation: "State" and "potential" are two aspects of a situation. From a presentation perspective, "state" refers to more tangible scenes, entities, events, environments; the "potential" refers to the magnitude and development direction of the potential influence formed by the "state" which can be obtained by understanding and analyzing.

And (3) situation estimation: the method is used for commanding decision-making personnel to sense situation elements in a certain time and space environment, and recognizing the acquired information so as to form prediction of the next moment of the situation elements.

Target basic information: the method mainly comprises strategic target basic information such as general profiles, position functions, composition distribution, structural properties, defense facilities and the like, and target basic information such as country, category, common residence (airport/harbor), membership, operational/reconnaissance/security capability and the like.

Target feature/rule information: the method mainly comprises target perception characteristic information such as visible light imaging characteristics, SAR imaging characteristics, infrared imaging characteristics, spectral characteristics, electromagnetic characteristics and the like, and target characteristic rule information such as an activity time rule, an activity path rule, a communication rule, a supply rule, a motion editing rule and the like; the target characteristic/rule information can provide support for finding target abnormity, analyzing target intention and predicting target movement.

Target grouping: also called target grouping or target aggregation, is an information fusion secondary stage and is an important function which needs to be realized by situation estimation.

According to the embodiment of the invention, the method for predicting the activity rule of the aerial target is provided, the reconnaissance/training/patrol rule of the target is extracted and analyzed based on the grasped historical track data of the enemy target, and as shown in fig. 1, the method comprises the following steps: :

step 1, pretreatment stage: screening the original data of investigation/training/patrol, and simultaneously carrying out normalization processing and formatting processing on the target track information to form a standard data format required by modeling.

Step 2, generating a preliminary track: and obtaining the segmented track according to the DBSCAN method for the track.

Step 3, extracting multi-dimensional characteristics of the track: updating the multidimensional characteristics of the track according to the height change, speed change, direction angle change, acceleration change and airspace change of the segmented track calculation target, and storing the multidimensional characteristics in a multidimensional track sample library;

step 4, constructing a Catboost classifier: dividing a training set, a verification set and a test set based on the accumulated data, generating a reconnaissance/training/patrol rule model, carrying out reconnaissance/training/patrol track updating, and storing the track into a route model library;

according to the embodiment of the invention, the analysis and mining of the activity rule needs to analyze historical data based on a mathematical statistics method, extract the frequently-occurring range of the target and the typical track of the target, and analyze the typical track of the target.

In step 2, generating a preliminary track: and obtaining the segmented track according to the DBSCAN method for the track. The track is extracted based on historical data, and a DBSCAN method is adopted for segmenting the track to obtain segmented track points.

And in the step 3, calculating the segmentation track points after the segmentation track points are obtained, and extracting the height change characteristic, the speed change characteristic and the direction angle change characteristic airspace change characteristic of the track.

In the step 3, the obtained multi-dimensional track data is divided into a training set, a verification set and a test set, and a CATBOOST classifier is adopted for training and predicting. And predicting the activity rule by adopting the classifier after the model parameters are optimized.

In the step 1, the original spatial data of the target object is filtered and the format is normalized. And then, grouping the normalized data according to the specified grouping quantity, respectively processing each group and selecting a core object to obtain a core object list of each group. And then, carrying out local clustering according to the core object list, and selecting a representative point of the local clustering. And finally, carrying out global confirmation to obtain clusters.

It should be noted here that the representative object cannot be selected too much or too little. If too much, the efficiency of the algorithm is difficult to exert, and the selection significance is lost; on the contrary, if the number of the objects is too small, the neighborhood of the selected representative object is difficult to compare and completely covers the neighborhoods of other objects, so that the objects are lost, one cluster is divided into a plurality of clusters, and the efficiency of the algorithm is difficult to develop. The improved DBSCAN algorithm comprises the following steps:

(1) given a set of data objects D and a data object p, p ∈ D, and defining the epsilon neighborhood N of p_ε(p) is N_εAnd (p) { q ∈ D | dist (p, q) ≦ epsilon }, wherein dist (p, q) represents the distance between two data objects p and q in D, and a parameter epsilon is dynamically set by using the non-space attribute of the data, so that epsilon becomes tv, wherein t is the set maneuvering time threshold value, and v is the maneuvering speed.

(2) The data set is scanned and an arbitrary data object p is selected. If p has been classified as a cluster or has been marked as noise, the data set is scanned again to select data objects. Otherwise, judging whether the point in the field is smaller than MinPts, if so, marking the point p as a boundary point or a noise point. Otherwise, marking the p point as a core point, establishing a new cluster C and adding all the p point fields into the C.

(3) For the problem that the core object is only qualified for expansion, once a certain core object is found, it needs to be determined whether each point in its neighborhood is a core object, so as to further expand outwards. If its neighborhood is a high density region, it means that there are many core objects, their neighborhoods overlap each other, and the neighborhoods of some points are even completely covered by the neighborhoods of other points, which causes repeated judgment, resulting in unnecessary consumption of time and memory. Therefore, it is necessary to select representative objects to reduce the clustering time and memory consumption.

Further optimization is possible for the above method.

Specifically, the main purpose of spatial clustering is to reduce the number of spatial target objects with excessive attention, so that the spatial target objects can be grouped into several classes without excessive attention. The target object should be divided into regions and grouped according to static and dynamic data of the position, then each group is clustered locally, and then the global judgment is performed to further reduce the memory.

According to the embodiment of the invention, the initial track points are given by sensors such as radar, and the track points are clustered by adopting an improved DBSCAN clustering method, so that the representative track points are selected to omit unimportant track points, and the algorithm efficiency is improved.

In said step 3, the changes of the characteristic quantities such as altitude, speed, direction angle, acceleration, etc. are all for a time interval t, and the values of these quantities at the previous moment are measured by various sensors, such as radar; the next moment in time can be measured again, and the difference between the two moments of time of the variable is the amount of change.

Aiming at uncertain factors such as nonlinearity, multivalue, repeatability, position deviation and the like existing in actual detection flight paths, a method which has good fault tolerance and can automatically distinguish high-value features is required to be adopted for classical flight path extraction. The Catboost realizes GBDT (gradient Boosting Decision Tree) efficiently, exerts speed and efficiency extremely, has excellent accuracy, and is widely concerned.

The Catboost is a GBDT framework with less parameters and high accuracy based on symmetric decision tree (oblique trees) implementation. The method can efficiently and accurately process the class type characteristics, and achieves the purposes of reducing overfitting and improving the accuracy and generalization capability of the algorithm by solving the problems of Gradient Bias (Gradient Bias) and Prediction shift (Prediction shift). Catboost embeds an innovative algorithm that automatically processes class features into numerical-type features. Firstly, statistics are carried out on category features, the frequency of occurrence of certain category features is calculated, and then a hyper-parameter is added to generate a new numerical feature. By using combined category features, feature dimensions can be enriched with relationships between features. The method of ordered boost is adopted to avoid the deviation of gradient estimation and solve the problem of prediction offset.

The category-type feature herein refers to not a numerical-type feature but a discrete-type feature, such as a place name of an airport in track prediction, such as the united states, or the like, or investigation, patrol, training, or the like in a category. In the gradient lifting algorithm, it is most common to convert these class features into numerical type for processing. A general class-type feature may be converted to one or more numerical-type features. If the cardinality of a certain type of feature is low (low-similarity features), that is, the number of elements in a set formed by removing all values of the feature is small, the feature is generally converted into a numerical type by using an One-hot encoding method. The One-hot coding can be completed during data preprocessing or model training, the latter method is more efficient to realize from the perspective of training time, and the latter method is also adopted by Catboost for class-type features with lower cardinality.

Among high cardinality class features (high cardinality features), such as userID, this encoding approach can generate a large number of new features, causing a dimensional disaster. One trade-off is that classes can be grouped into a limited population and then One-hot encoded. One common solution to the above method is to group according to Target variable Statistics (TS) which are used to estimate the Target variable expectation for each class. Even directly using the target variable statistics as a new numerical variable to replace the original category variable. By setting a threshold of the target variable statistic type feature, based on the logarithmic loss, the Kernel coefficient or the mean square error, the optimal one of all possible partitions dividing the class into two for the training set is obtained, namely for different loss functions, the optimal solution can be obtained by using the target variable statistic data. In LightGBM, the class-type feature is represented by Gradient Statistics (GS) at each step of Gradient boosting. Although providing important information for tree building, this approach has the disadvantage of increasing computation time because gradient statistics need to be computed at each step of the iteration for each class-type feature; the storage requirement is increased, and for a class type variable, the class of each node is required to be stored in each separation. LightGBM classifies all long-tailed classes into One class at the expense of losing part of the information, which is still better than One-hot coding when dealing with high-radix class-type features. Also for using the TS feature, only one number is calculated and stored for each category. The use of TS as a new numerical feature is the most efficient and information loss-minimizing method of processing the class-type features.

The design of the CatBOost algorithm was designed to better handle the categorical features in the GBDT feature. When dealing with the cationic features in the GBDT features, the simplest approach is to replace the average value of the tags corresponding to the cationic features. In the decision tree, the label mean will be the criterion for node splitting. This method is called Greedy Target-based Statistics, Greedy TS for short, and is expressed by the following formula:

where x, y, k, i, x_j,kAlpha and I respectively represent the average value of the label, the target value, the number of the training samples, the class characteristic and the ith characteristic of the kth training sample, alpha is a parameter larger than 0, P is the average value of target variables in all data, and I is an indicating function, namely when x is_j,k＝x_i,kIts value is taken to be 1 if not 0.

This approach has the obvious drawback that features typically contain more information than labels, and if one were to force the features to be represented by the average of the labels, condition drift issues arise when the training data set and the test data set differ in data structure and distribution. One standard way to improve Greedy TS is to add a priori distribution terms, which can reduce the impact of noise and low frequency class data on the data distribution.

CatBOost processes category features Category features statistics of some data are calculated. And calculating the occurrence frequency of certain category, adding a super parameter to generate new data feature numerical features. This strategy requires that the same label data cannot be arranged together, requiring a data set to be shuffled prior to training. Using different permutations of data, a round of dice is thrown to determine which permutation to use to generate a tree before each round of tree building. Different combinations of class features are contemplated. CatBoost considers only a portion of binding partners when the class characteristics of the desired combinations vary. When selecting the first node, only one feature, such as A, is considered to be selected, and when generating the second node, the best one is selected by considering the combination of A and any one category feature. The binding partners are thus generated using a greedy algorithm.

The data center has historical track information provided by each radar station, and as described in the step 3, in order to reduce the scale of data processing and increase the speed of data processing, the invention adopts an improved DBSCAN clustering algorithm to extract representative track points. Therefore, a representative course point is obtained.

For the track points, the invention innovatively adopts the CATBOOST method to carry out classification training, and representative data obtained by the improved DBSCAN method needs to be divided into a training set and a verification set test set for training. And training parameters by using a training set, and adjusting the hyper-parameter test set by using a verification set to test the performance of the algorithm, thereby obtaining a mathematical model for classifying the flight path data. Since the type information activity information of the aircraft all have discrete values, the CATBOOST algorithm does not need to convert these values into numbers in the data preprocessing stage. This is a gradient boosting algorithm that can boost the class characteristics well. Thus, the catbios algorithm is used to classify data tables consisting of position information, type information (fighters, scouts, drones, bombers, etc.) and activity information (scouts, trains, and patrols) in the flight path data.

Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, but various changes may be apparent to those skilled in the art, and it is intended that all inventive concepts utilizing the inventive concepts set forth herein be protected without departing from the spirit and scope of the present invention as defined and limited by the appended claims.

Claims

1. A method for predicting an activity rule of an aerial target is characterized by comprising the following steps:

2. The method according to claim 1, wherein in step 2, the data in the post-standard data format of the target object is grouped according to the designated number of groups and each group is processed respectively to select the core object, so as to obtain the core object list of each group, then, local clustering is performed according to the core object list, representative points of the local clustering are selected, and finally, global confirmation is performed to obtain the clustering.

3. The method for predicting the activity rule of the target object in the air according to claim 1, wherein the improved DBSCAN algorithm in the step 3 comprises the following steps:

(2) scanning the data set, selecting any data object p, and if p is classified into a certain cluster or marked as noise, scanning the data set again to select the data object; otherwise, judging whether the point in the field is smaller than MinPts, if so, marking the point p as a boundary point or a noise point; otherwise, marking the p point as a core point, establishing a new cluster C and adding all the p point fields into the new cluster C;

4. The method of claim 3, wherein the target activity rule prediction method,

specifically, the main purpose of spatial clustering is to reduce the number of spatial target objects with excessive attention, and further optimize the number as follows: the method comprises the steps of firstly carrying out region division grouping on target objects according to static and dynamic data of positions, then carrying out local clustering on each group, and then judging from the whole situation so as to further reduce the memory.

5. The method according to claim 1, wherein in step 3, the changes of the characteristic quantities such as altitude, speed, direction angle, acceleration, etc. are measured by various sensors at the previous moment and measured again at the next moment, and the difference between the two moments of the variables is the change of the characteristic.

6. The method for predicting the activity rule of the aerial target according to claim 1, wherein the step 4 is a classification training by using a CATBOOST method, and comprises the steps of firstly, dividing representative data obtained by an improved DBSCAN method into a training set and a verification set test set; training parameters by using a training set, and adjusting a hyper-parametric test set by using a verification set to test the performance of the algorithm to obtain a mathematical model for classifying track data; and classifying a data table consisting of position information, type information and activity information in the flight path data by adopting a CATBOOST algorithm, and finally predicting the target activity rule by utilizing a trained model.