CN-121999431-A - Target maneuvering unit identification method based on self-attention convolutional network

CN121999431ACN 121999431 ACN121999431 ACN 121999431ACN-121999431-A

Abstract

The invention discloses a target maneuvering unit identification method based on a Self-care convolution network, which comprises the steps of obtaining historical time track data of a target aircraft, extracting a plurality of instantaneous features, judging maneuvering actions according to the plurality of instantaneous features, judging according to accumulated features if the same maneuvering actions are executed in a certain time period, dividing the track into maneuvering tracks of a plurality of sections of corresponding maneuvering unit categories, constructing a sample set, dividing a training set, generating a Self-CovTabNet network, training the Self-CovTabNet network by using the training set to obtain an identification network, and inputting the instantaneous features extracted from the current time track data of the target aircraft into the identification network to obtain the maneuvering unit category to which the maneuvering track of the current time track belongs. The invention can realize high-precision and high-efficiency motorized unit identification.

Inventors

WANG YUAN
LIU SUYI
ZHANG SHUMING
REN FENG
MA SHICHAO
LIAN XIAOBIN
REN MINGXUAN
NING XIN
Cao Xiuyang
WANG YUXIN

Assignees

西北工业大学

Dates

Publication Date: 20260508
Application Date: 20251226

Claims (7)

1. A target maneuvering unit identification method based on a self-attention convolution network is characterized by comprising the following steps: Step 1, acquiring track data of a target aircraft in historical time, extracting a plurality of instantaneous characteristics of the target aircraft at each moment in the historical time according to the track data, and judging and obtaining maneuvering actions executed by the target aircraft at each moment in the historical time according to the extracted plurality of instantaneous characteristics of the target aircraft at each moment in the historical time; if the accumulated characteristics meet the requirements, the maneuvering tracks formed by the target aircraft executing the same maneuvering actions in the corresponding time period are divided into corresponding maneuvering unit categories according to the maneuvering actions corresponding to the corresponding time period, so that the tracks of the target aircraft in the historical time are divided into maneuvering tracks of a plurality of sections of corresponding maneuvering unit categories, wherein the maneuvering unit categories comprise a flat flight category, a horizontal left turn category, a horizontal right turn category under a horizontal plane maneuvering unit, a vertical concave climbing category, a vertical concave diving category under a vertical plane maneuvering unit, a left turn concave climbing category, a left turn convex diving category under a space maneuvering unit, a right turn concave climbing category, a right turn concave diving category, a left turn convex diving category, a right turn concave diving category and a right turn convex diving category under a space maneuvering unit; Constructing samples according to the instant characteristics of each time corresponding to each section of maneuvering track obtained by dividing, so as to obtain a plurality of samples, wherein each sample is provided with maneuvering unit class labels to which the corresponding maneuvering track belongs, a sample set is formed by all samples, and a training set and a verification set are divided from the sample set; step 2, generating a Self-CovTabNet network, wherein the Self-CovTabNet network comprises an identification part and an output part, and the identification part comprises a plurality of parallel links and a weight merging module; In the identification part, each link comprises a plurality of layers, and each layer comprises a self-attention transducer module and a convolution characteristic transducer module; The output of the self-attention transducer module in each layer of each link is multiplied with the original data and then used as the input of the convolution characteristic transducer module in the corresponding layer; The output of the convolution characteristic transducer module in each layer after the first layer of each link is connected with the output of the convolution characteristic transducer module in the adjacent upper layer in a residual way and then is used as the input of the self-attention transducer module in the adjacent lower layer; The output of the last layer of convolution characteristic transducer module in each link is subjected to residual connection with the output of the adjacent last layer of convolution characteristic transducer module, and then weight combination is carried out according to decision weights of each link through a weight combination module to be used as the output of the identification part; The output part of the Self-CovTabNet network adopts a Softmax function, and the result output by the identification part is processed by the Softmax function to obtain the prediction probabilities of different categories; Training the Self-CovTabNet network by adopting the training set in the step 1, inputting each sample in the training set to a Self-attention transducer module of a first layer of each link in a Self-CovTabNet network identification part during training, and taking the trained Self-CovTabNet network as an identification network; And 3, extracting instantaneous characteristics of each moment from the current time track data of the target aircraft, inputting the instantaneous characteristics of each moment into the identification network, and identifying by the identification network to obtain the maneuvering unit category to which the maneuvering track corresponding to the current time track of the target aircraft belongs.
2. The method for identifying a target maneuver unit based on the self-care convolutional network as recited in claim 1 wherein in step 1, the plurality of transient features includes a rate of change of the track yaw angle at each time instant Track inclination angle Track pitch rate of change The accumulated characteristic is the accumulated track dip angle change rate of the time period Or accumulated track deflection angle change rate 。
3. The method for identifying a target maneuver unit based on a self-care convolutional network as recited in claim 2 wherein in step1, the rate of change of the track yaw angle at each time is combined Track inclination angle Track pitch rate of change Judging the maneuver executed by the target aircraft at the corresponding moment; If the maneuver at each moment in a certain time period is the same, extracting the accumulated track dip angle change rate in the corresponding time period Or accumulated track deflection angle change rate And judging the change rate of the accumulated track dip angle Or accumulated track yaw rate Or accumulated track pitch rate And accumulated track deflection angle change rate If so, dividing the maneuver tracks formed by the target aircraft executing the same maneuver in the corresponding time period into corresponding maneuver unit categories according to the maneuver actions corresponding to the corresponding time period.
4. The method for identifying a target mobile unit based on a self-care convolutional network according to claim 1, wherein in the step 2, the self-care transform module comprises three groups of linear change layers, point integration evaluation layers, softmax function layers, multiplication modules, GBN layers and Entmax layers; the convolution feature transducer module comprises a full connection layer, a stacked multi-layer one-dimensional convolution, a GBN layer, a Dropout layer and a DE-GLU layer.
5. The method for identifying a target mobile unit based on a Self-care convolutional network as recited in claim 4, wherein in step 2, the data processing procedure of Self-CovTabNet network during training is as follows: Combining a plurality of instantaneous features of each sample in the training set into feature vectors according to a preset sequence, and mapping the feature vectors through an embedding layer to obtain an embedding matrix corresponding to each sample; generating a position information vector for representing the serial number of each instantaneous feature in the feature vector, and adding an embedded matrix of each sample with the position information vector to obtain each sample after position coding, wherein the samples after position coding are input to a Self-attention transducer module of a first layer in each link of a Self-CovTabNet network identification part; In each link, the self-attention transducer module of the first layer obtains a query Q, a key K and a value V of each sample based on the sample representation after position coding through three groups of linear transformation layers, calculates a scaling dot product score between the query Q and the key K corresponding to each sample through a dot product score layer, calculates self-attention weight distribution of each sample based on the dot product score through a Softmax function layer, combines the self-attention weight corresponding to each sample with the corresponding value V through weighted summation to obtain corresponding first-layer self-attention features, normalizes the corresponding first-layer self-attention features of each sample through a GBN layer, performs sparse mapping on the normalized first-layer self-attention features through a Entmax layer, and outputs a first-layer feature selection mask M 1 , wherein the mask M 1 is a sparse feature selection weight vector, and is used for performing weighted screening on input instantaneous features so as to inhibit irrelevant features and highlight key features, so as to obtain local feature representations for subsequent convolution feature extraction; in each link, the first layer characteristic selection mask M 1 output by the first layer self-attention transducer module is multiplied by the embedding matrix of the corresponding sample element by element or by dimension weighting so as to extract the first layer local characteristic from the sample; In each link, a convolution feature transducer module of a first layer carries out linear mapping and feature fusion on a first layer local feature of each sample to obtain a first layer high-dimensional potential space representation of the sample, extracts a first layer global feature from the first layer potential space of each sample through stacked multi-layer one-dimensional convolution to obtain each sample with the first layer local feature and the first layer global feature, normalizes each sample with the first layer local feature and the first layer global feature by using a GBN layer and outputs the normalized sample to a Dropout layer, the Dropout layer carries out random inactivation on feature dimensions, and outputs the Dropout layer to a DE-GLU layer; The method comprises the steps of obtaining a first layer of self-attention weight distribution of each sample, obtaining a second layer of self-attention feature by combining the self-attention weight of each sample with a corresponding value V through weighted summation, normalizing and outputting the second layer of self-attention feature corresponding to each sample to a Entmax layer through a GBN layer, and finally sparsifying and mapping the normalized second layer of self-attention feature through a Entmax layer, and outputting a second layer of feature selection mask M 2 , wherein the feature selection mask M 2 is a sparse weight vector corresponding to an input feature dimension; In each link, the second layer characteristic selection mask M 2 output by the second layer self-attention transducer module is multiplied by the embedding matrix of the corresponding sample element by element or by dimension weighting so as to extract the second layer local characteristic from the sample; In each link, a convolution feature transducer module of a second layer firstly carries out linear mapping/feature fusion on a second layer local feature of each sample through a full connection layer to obtain a second layer high-dimensional potential space of the sample, then extracts a second layer global feature from the second layer potential space of each sample through stacked multi-layer one-dimensional convolution to obtain each sample with the second layer local feature and the second layer global feature, normalizes each sample with the second layer local feature and the second layer global feature by using a GBN layer and then outputs the normalized sample to a Dropout layer, the Dropout layer carries out random inactivation on feature dimensions, and outputs the Dropout layer to a DE-GLU layer; In each link, each sample with the second layer local feature and the screened second layer global feature output by the second layer convolution feature transducer module is connected with each sample with the first layer local feature and the screened first layer global feature output by the first layer convolution feature transducer module in a residual way, so that each sample after residual connection is obtained and used as a decision result of the second layer in each link; In each link, the decision result of the M-1 th layer after the first layer is output to a self-attention transducer module of the M layer, wherein M is more than or equal to 3, the self-attention transducer module of the M layer firstly obtains the query Q, the key K and the value V of each sample based on the representation and the position information of the sample corresponding to the decision result of the M-1 th layer through three groups of linear transformation layers, then calculates the scaling dot product score between the query Q and the key K corresponding to each sample through a dot product score layer, calculates the self-attention weight distribution of each sample based on the dot product score through a Softmax function layer, combines the self-attention weight corresponding to each sample with the corresponding value V to obtain the corresponding self-attention characteristic of the M layer through weighted summation, then normalizes and outputs the self-attention characteristic of the M layer corresponding to the Entmax layer through a GBN layer, finally sparsely maps the normalized self-attention characteristic of the M layer through a Entmax th layer, and outputs a mask M m , wherein the feature selection mask M m is a mask vector corresponding to the sparse dimension feature; in each link, the M-layer feature selection mask M m output by the M-layer self-attention transducer module is multiplied by the embedding matrix of the corresponding sample element by element or by dimension weighting to extract the M-layer local feature from the sample, and the sample with the M-layer local feature is input to the M-layer convolution feature transducer module In each link, an mth layer convolution feature transducer module firstly carries out linear mapping/feature fusion on a second layer local feature of each sample through a full connection layer to obtain an mth layer high-dimensional potential space of the sample, then extracts a second layer global feature from the mth layer potential space of each sample through stacked multi-layer one-dimensional convolution to obtain each sample with mth layer local feature and mth layer global feature, then normalizes each sample with mth layer local feature and mth layer global feature by using a GBN layer and outputs the normalized sample to a Dropout layer, the Dropout layer carries out random inactivation on feature dimension, and outputs the Dropout layer to a DE-GLU layer; In each link, each sample with the m-th local feature and the screened m-th global feature output by the m-th convolution feature transducer module is connected with each sample with the m-1-th local feature and the screened m-1-th global feature output by the m-1-th convolution feature transducer module in a residual way, each sample after residual connection is obtained, and the sample is used as a decision result of the m-th layer in each link and is output to a self-attention transducer module in the m+1-th layer; Finally, each sample after residual connection in the decision result obtained by the last layer in each link is subjected to weight combination according to the decision weight of each link by a weight combination module and is used as the output of the identification part; and processing the result output by the identification part through a Softmax function to obtain the prediction probability of each sample belonging to different maneuvering unit categories.
6. The method for identifying the target maneuvering units based on the Self-attention convolution network according to claim 5, wherein a loss function adopted in training is cross entropy loss, and the cross entropy loss is obtained by calculating maneuvering unit category prediction probability and corresponding real category labels output by the Self-CovTabNet network on each sample in a training set.
7. The method for identifying the target maneuvering unit based on the Self-care convolutional network according to claim 5, wherein during training, samples in a training set are sent to the Self-CovTabNet network in batches for training, parameters of the Self-CovTabNet network are updated by back propagation based on a loss function calculation result after each batch training is finished, whether the Self-CovTabNet network is converged or stopped early after the current batch training is verified through a verification set, parameters of the Self-CovTabNet network after the current batch training are cured if the Self-CovTabNet network is converged or stopped early, a trained Self-CovTabNet network is obtained, and samples in the next batch are sent to the Self-CovTabNet network for training until the Self-CovTabNet network is converged or stopped early after the training, and the trained Self-CovTabNet network is obtained if the Self-CovTabNet network after the current batch training is not converged or stopped early.

Description

Target maneuvering unit identification method based on self-attention convolutional network Technical Field The invention relates to the field of aircraft maneuvering track recognition methods, in particular to a target maneuvering unit recognition method based on a self-attention convolution network. Background In complex dynamic scenes under over-the-horizon conditions, the target aircraft is usually continuously maneuvered over a wide airspace, with a strong implication of its intent to move. Because the scene has the characteristics of high non-collaboration and dynamic evolution, in order to realize accurate perception of the motion state of the target aircraft and effective scheduling of system resources, the continuous maneuvering track of the target aircraft is subjected to fine modeling, divided into maneuvering unit categories with clear physical meanings, and maneuvering intention recognition and intelligent decision are developed on the basis. The maneuvering unit category identification of the maneuvering track is a core link for connecting the original track with the high-level intention information, and the identification effect directly influences the accuracy and the robustness of the subsequent state estimation, the behavior inference and the task planning. Therefore, facing the above-mentioned complex non-cooperative dynamic scenario, it is highly desirable to realize real-time and accurate identification of the target maneuvering unit, and provide reliable basic input for constructing a high-performance situation awareness and intelligent decision system. In the prior art, researches on the category identification of maneuvering units to which maneuvering trajectories of target aircrafts belong are mainly focused on three kinds of ideas. The method is based on expert experience and maneuvering mechanism analysis, and is characterized in that the track is divided into a plurality of categories such as flat flight, turning, climbing, diving and the like by setting thresholds and rules for indexes such as track angle, entry angle change and overload, the method is simple to realize and high in interpretation, but depends on manual setting rules seriously, parameters are required to be adjusted frequently under different task backgrounds and sensing conditions, and the method is difficult to adapt to complex and changeable and high-noise environments. The second type of method is mainly based on traditional machine learning, the manually constructed geometric and kinematic features are input into a support vector machine, a decision tree, a random forest and the like model, the class of the maneuvering unit is automatically judged, and compared with a pure rule method, the method can more fully utilize data, but is highly dependent on feature engineering, the high-order nonlinear relation among features and the local time sequence mode are limited in depicting capability, and when sample distribution changes or complex combination maneuvering occurs, the identification performance is easy to be obviously reduced. The third type of method is mainly based on deep learning thought, utilizes structures such as a multi-layer perceptron, a convolution network or a circulation network to classify maneuvering fragments, alleviates the problem of difficult design of artificial features to a certain extent, but mostly directly uses a general network in an image or sequence task, does not carry out structural optimization on the data characteristics of 'table type, multi-feature and high coupling' in maneuvering unit category identification, and is easy to cause the problems of insufficient attention to key features, insufficient suppression of noise and redundant features and the like because feature selection and feature extraction are often interwoven in the same network. In addition, in the currently published patent and literature, the task of adopting TabNet and variants thereof for motor unit class identification is still less, and the existing method still has room for improvement in aspects of feature selection, global and local feature fusion, stable training of deep structures and the like. Therefore, in the non-collaborative target behavior analysis under the complex dynamic scene, a new method for identifying the maneuvering unit category, which can perform self-adaptive feature selection and depth characterization and give consideration to identification precision and engineering realizability to the tabular maneuvering features, is needed, so that more reliable basic input is provided for the following various maneuvering intention identification and intelligent decision tasks. Disclosure of Invention The invention provides a target maneuvering unit identification method based on a self-attention convolution network, which is oriented to analysis and modeling requirements of non-cooperative target maneuvering behaviors in a complex dynamic scene, and aims to realize high-precision and hi