CN-121524744-B - Target identification method based on unsupervised dynamic-attribute feature decoupling network

CN121524744BCN 121524744 BCN121524744 BCN 121524744BCN-121524744-B

Abstract

The invention provides a target identification method based on an unsupervised dynamic-attribute feature decoupling network, which comprises the steps of obtaining HRRP data of a target in a period of time as an HRRP sequence, processing the target HRRP sequence by using a trained HRRP target identification model to obtain a target identification result, wherein the HRRP target identification model comprises the unsupervised dynamic-attribute feature decoupling network and the HRRP target identification network, obtaining attribute features after the obtained target HRRP sequence is processed by the trained unsupervised dynamic-attribute feature decoupling network, and processing the attribute features by the trained HRRP target identification network to obtain the target identification result. The HRRP target identification method provided by the invention can more accurately and robustly identify the type of the target under the condition that various dynamic factors such as azimuth change, noise interference and the like are overlapped and the numerical value is unknown.

Inventors

XU YUELEI
ZHANG FAN
ZHANG ZHAOXIANG
WANG XUAN
WANG HONGQIAO
WANG XUANBIN
CHEN HAN

Assignees

西北工业大学

Dates

Publication Date: 20260508
Application Date: 20251117

Claims (10)

1. A target identification method based on an unsupervised dynamic-attribute feature decoupling network is characterized by acquiring HRRP data of a target within a period of time as an HRRP sequence, and processing the target HRRP sequence by using a trained HRRP target identification model to obtain a target identification result; The HRRP target recognition model comprises an unsupervised dynamic-attribute feature decoupling network and an HRRP target recognition network, wherein the obtained target HRRP sequence is processed by the trained unsupervised dynamic-attribute feature decoupling network to obtain attribute features, and the trained HRRP target recognition network processes the attribute features to obtain a target recognition result; The unsupervised dynamic-attribute feature decoupling network comprises a dynamic feature encoder, an attribute feature encoder and a decoder, wherein the unsupervised dynamic-attribute feature decoupling network adopts an asymmetric variation reasoning architecture, the dynamic feature encoder and the attribute feature encoder respectively model posterior distribution of dynamic features and attribute features, and the decoder is used for modeling conditional likelihood distribution of HRRP samples.
2. The method for identifying the target based on the unsupervised dynamic-attribute feature decoupling network of claim 1, wherein the method is characterized by comprising the following steps: The dynamic feature encoder consists of an LSTM layer and a plurality of FC layers, and is input with an HRRP sequence Average division into After a number of non-overlapping windows, Data for each window And corresponding mask Firstly, inputting an LSTM layer for recursion processing, capturing dynamic dependency relationship among time steps, and outputting the LSTM layer Mapping to Gao Weiqian space via FC layer to obtain features Respectively estimating the mean value of the potential characteristic distribution through two independent FC layers Sum covariance parameter Will be The mean and covariance parameters of the non-overlapping windows are stacked to obtain the overall mean Sum covariance parameter Using the mean of the whole Sum covariance parameter Constructing posterior distribution of dynamic features Posterior distribution of dynamic features Sampling to obtain Non-overlapping windows Corresponding dynamic characteristics ; The attribute feature encoder consists of an LSTM layer and an FC layer, an input HRRP sequence Random interception and continuous Sub-sequence of individual time steps And corresponding mask First inputting LSTM layer for recursion treatment, outputting LSTM layer Mapping to Gao Weiqian space via FC layer to obtain features Obtaining the mean value estimation of the characteristic distribution through the FC layer Establishing the mean value as Normal distribution with standard deviation of 1 as posterior distribution of attribute characteristics Posterior distribution of attribute features Sampling to obtain a sequence Attribute features of (a) ; The input of the decoder being a dynamic feature And attribute features Is a fusion feature of (2) The fusion feature After being processed by the FC layer and the LSTM layer, the average value of the generated sample distribution is estimated by two independent FC layers And standard deviation By means of the mean value And standard deviation Conditional likelihood distribution for obtaining HRRP sequence samples 。
3. The method for identifying the object based on the unsupervised dynamic-attribute feature decoupling network of claim 2, wherein the method is characterized by using the average value of the whole Sum covariance matrix Constructing posterior distribution of dynamic features Wherein the covariance matrix Calculated using the following formula: Wherein the method comprises the steps of Is the noise added on the diagonal line and, To utilize covariance parameters And (5) constructing a sparse matrix.
4. The method for identifying an object based on an unsupervised dynamic-attribute feature decoupling network of claim 1, wherein said HRRP object identification network is characterized by attributes As an input, discrimination of the target class is performed by a classifier composed of a plurality of FC layers.
5. The method for identifying the target based on the unsupervised dynamic-attribute feature decoupling network of claim 2, wherein the method is characterized in that the unsupervised dynamic-attribute feature decoupling network is trained in an unsupervised manner by utilizing multi-element HRRP sequence sample data, and comprises the following specific processes: (1) Randomly initializing parameters of an unsupervised dynamic-attribute feature decoupling network ; (2) Training data from multiple HRRP sequence samples A batch of data is randomly sampled and sent into an unsupervised dynamic-attribute characteristic decoupling network; (3) Forward propagation of the network and calculation of the total loss function value Including evidence lower bound loss And the inverse regularization loss : Wherein the method comprises the steps of The weight of the loss is regularized for the counterfactual, and the overall loss is minimized in the network training process; Evidential lower bound loss According to the formula Calculated, wherein Is the batch size of HRRP sequence samples, Is a negative log-likelihood term that is used to determine the likelihood, As the weight of the material to be weighed, To measure dynamic characteristic posterior distribution With its a priori distribution The degree of divergence of the differences between them, Posterior distribution for metrology attribute features With its a priori distribution Divergence of the differences between; Counterfactual regularization loss According to the formula The calculation result shows that the method comprises the steps of, Is attribute characteristics And anti-facts attribute feature In the case of a counterfactual sample Posterior probability difference below, wherein the prior distribution of the attribute features Random sampling to construct a counterfactual attribute feature Dynamic characteristics Anti-facts attribute feature Input decoder for generating a counterfactual intervention sample ; (4) Updating parameters of an unsupervised dynamic-attribute feature decoupling network by a back propagation algorithm ; (5) Repeating the steps (2) - (4), and completing the unsupervised dynamic-attribute feature decoupling network training after the set iteration termination condition is reached.
6. The method for identifying a target based on an unsupervised dynamic-attribute feature decoupling network of claim 5, wherein the method is characterized by modeling a priori distribution of dynamic features using a Gaussian process Using standard normal distribution Prior distribution as attribute features 。
7. The method for identifying a target based on an unsupervised dynamic-attribute feature decoupling network of claim 5, wherein said plurality of HRRP sequence sample data comprises HRRP data having different degrees of azimuthal absence.
8. The method for identifying the target based on the unsupervised dynamic-attribute feature decoupling network of claim 2, wherein the training of the HRRP target identification network by using the plurality of HRRP sequence sample data comprises the following steps: (1) Freezing parameters of an unsupervised dynamic-attribute feature decoupling network, randomly initializing parameters of an HRRP target recognition network ; (2) Training data from multiple HRRP sequence samples The method comprises the steps that attribute characteristics obtained after a batch of data randomly sampled is processed by an unsupervised dynamic-attribute characteristic decoupling network and corresponding target class labels are input into an HRRP target identification network; (3) Forward propagation by the network and calculating the cross entropy loss value: Wherein, the Is the batch size of HRRP sequence samples, Is the number of sample classes that are to be counted, Is the first The true class of the individual samples is that, The method is a class probability distribution output by an HRRP target recognition network, and aims at minimizing the cross entropy function in the training process; (4) Updating parameters of HRRP target identification network by back propagation algorithm ; (5) Repeating the steps (2) - (4), and completing HRRP target recognition network training after the set iteration termination condition is reached.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the object recognition method according to any one of claims 1 to 8 when executing the program.
10. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the object recognition method according to any one of claims 1 to 8.

Description

Target identification method based on unsupervised dynamic-attribute feature decoupling network Technical Field The invention relates to the technical field of target identification, in particular to a target identification method based on an unsupervised dynamic-attribute feature decoupling network. Background In the scenes of offshore law enforcement, maritime search and rescue and the like, the searching radar is utilized for searching and identifying targets, which is a main means in severe weather environments. The radar high-resolution range profile (High Resolution Range Profile, HRRP) contains distinguishing characteristics such as material properties, geometric structures and the like of the targets, so that the radar high-resolution range profile is widely used for radar automatic target recognition and is helpful for quickly finding the targets. HRRP consists of the magnitude of the complex echo vector sum of the target scatterers in each range bin, which represents the projection of the complex echo of the target scattering center onto the radar Line-of-Sight (LOS). In recent years, due to the development of deep learning technology, research on HRRP target identification based on deep neural network has been significantly advanced, but there are a plurality of difficulties and challenges, including the problem of orientation sensitivity of HRRP. Because the change of the relative azimuth of the radar and the target can cause the projection distribution of the scattering points to change, the HRRP waveforms presented by the same target under different observation azimuth have obvious differences, and the identification model is easy to identify by mistake under the condition that the waveforms are not seen. In many existing HRRP target recognition methods, a recognition model is usually constructed under the assumption of a complete azimuth sample, and the assumption model sees HRRP data of all azimuths of the target, so that the method has good recognition performance. However, in an actual non-cooperative target identification scene, the available HRRP data often has serious azimuth loss due to the uncontrollability of target movement and the constraint of observation conditions. Therefore, how to realize the comprehensive domain generalization identification based on partial azimuth samples has very important practical significance. Related scholars have conducted research around this problem, and can be mainly divided into two directions of data enhancement and model enhancement. Data enhancement is mainly by generating HRRP data against network or variational coded structure synthetic missing orientations, but the physical rationality and distribution consistency of the generated samples lacks effective evaluation criteria. Model enhancement directions comprise meta-learning cross-azimuth adaptation methods, migration feature mapping models, self-supervision comparison frameworks and the like, and the methods are good in performance under the condition of medium loss rate, but are remarkably reduced in performance under the condition of high-proportion azimuth loss. It is worth noting that, inspired by decoupling representation learning, a learner puts forward a type-azimuth decoupling network, category and azimuth characteristics of targets are separated through double-branch supervised learning, then target identification is carried out by utilizing the separated category characteristics, and a new thought is provided for HRRP target identification under azimuth deletion. However, this method relies on dual supervisory signals, requiring simultaneous acquisition of the target class tag and accurate bearing information. The azimuth information of the target is determined by the LOS direction of the radar and the gesture of the target, but in an actual scene, the gesture of the target is complex and changeable due to environmental factors, and even if the LOS direction of the radar relative to the target point can be acquired, the gesture data of the target body is still difficult to accurately acquire. Thus, in such cases, the position information of the target is difficult to accurately provide, resulting in the existing HRRP decoupling network facing applicability challenges. In addition, factors such as noise interference, radar hardware faults, unstable data transmission and the like which possibly change at any time can also cause the change of HRRP waveform, when such time-varying unknown interference factors exist in the environment, the specific values of the interference factors are difficult to obtain, the interference factors cannot be effectively stripped by the existing decoupling network based on double supervision, and the network cannot extract pure target class characteristics, so that the final recognition performance is affected. Therefore, there is an urgent need for a more flexible and robust HRRP decoupling strategy to extract target intrinsic attribute features, ther