CN-121978650-A - Radar HRRP deformation target identification method based on feature correction

CN121978650ACN 121978650 ACN121978650 ACN 121978650ACN-121978650-A

Abstract

The invention discloses a radar HRRP deformation target identification method based on feature correction, which comprises the steps of preprocessing radar one-dimensional HRRP echo data of a target to be identified to obtain HRRP original target data, superposing block sparse signals with different sparsity and Gaussian white noise on the HRRP original target data to obtain HRRP deformation target data, constructing a lightweight feature extraction network, performing pretraining by using the HRRP original target data, constructing a feature correction network comprising a multi-dimensional parallel feature correction module, a feature fusion module and a denoising post-processing module, performing training after cascading with the feature extraction network, and combining the feature extraction network and the feature correction network to obtain a deformation target identification model so as to realize identification and classification of deformation targets. The method can effectively recover the intrinsic characteristic distribution of the deformed target, and remarkably improves the recognition accuracy and generalization capability of the non-cooperative target in the low signal-to-noise ratio environment.

Inventors

WANG PENGHUI
LIU HONGWEI
SONG BIN
GUO SHUAI
CHEN BO
CHEN WENCHAO
DING JUN

Assignees

西安电子科技大学

Dates

Publication Date: 20260505
Application Date: 20260227

Claims (10)

1. The radar HRRP deformation target identification method based on the feature correction is characterized by comprising the following steps of: S1, acquiring radar one-dimensional HRRP echo data of a target to be identified, preprocessing the radar one-dimensional HRRP echo data, and taking the preprocessed data as HRRP original target data; S2, overlapping block sparse signals with different sparsity on the HRRP original target data to serve as deformation components, and further overlapping Gaussian white noise to obtain HRRP deformed target data; correlating the HRRP deformation target data, the HRRP original target data and the corresponding target class labels to construct a pair of sample sets for subsequent network training; S3, constructing a lightweight feature extraction network, and pre-training the feature extraction network by utilizing the HRRP original target data to obtain a pre-trained feature extraction network; S4, constructing a characteristic correction network comprising a multidimensional parallel characteristic correction module, a characteristic fusion module based on a multi-branch attention mechanism and a denoising post-processing module based on an improved Pureformer network, cascading the characteristic correction network with the characteristic extraction network, fixing parameters of the characteristic extraction network, taking the paired sample sets as input data, and optimizing the parameters of the characteristic correction network through joint training to obtain a trained characteristic correction network; and S5, combining the feature extraction network obtained by the pre-training in the step S3 with the feature correction network obtained by the training in the step S4 to obtain a deformation target recognition model based on feature correction and Pureformer post-processing so as to realize recognition and classification of the deformation target by using the model.
2. The method for identifying a radar HRRP deformation target based on feature correction according to claim 1 wherein in S3, the feature extraction network adopts an ultra-lightweight residual compression-excitation network structure, or adopts a network structure based on ResNet, denseNet, or adopts a network structure based on channel attention and coordinate attention.
3. The method for identifying the radar HRRP deformation target based on feature correction of claim 1, wherein in S3, the feature extraction network adopts an ultra-lightweight residual compression-excitation network structure, including an input layer, an enhanced residual layer and an output layer, and the pre-training the feature extraction network by using the HRRP original target data comprises: s31, inputting the HRRP original target data into an input layer of the feature extraction network, and sequentially performing preliminary feature extraction and normalization processing including batch normalization, nonlinear activation and random inactivation regularization to obtain a first feature; S32, inputting the first features into an enhanced residual layer of the feature extraction network, and sequentially performing multi-layer convolution processing, feature enhancement processing based on a channel attention weighting mechanism and residual connection fusion processing to obtain second features; s33, inputting the second features to an output layer of the feature extraction network, and performing feature aggregation and mapping to generate primary feature representation; and S34, back-propagating the feature extraction network by using a cross entropy loss function, and optimizing parameters of the iterative feature extraction network until the pre-training is completed.
4. The method for identifying the radar HRRP deformation target based on the feature correction according to claim 1, wherein in S4, the multi-dimensional parallel feature correction module adopts a parallel spatial correction and channel correction structure, or a multi-dimensional self-attention mechanism and a non-local attention mechanism, or a feature relation modeling mode based on a graph neural network, and performs correlation modeling and correction on a feature representation output by the feature extraction network to obtain multi-dimensional correction features; the characteristic fusion module based on the multi-branch attention mechanism adopts a three-branch attention fusion network structure to fuse multidimensional correction characteristics so as to obtain fusion characteristics; The denoising post-processing module based on the improved Pureformer network comprises a feature enhancement module, a multi-head depth separable self-attention module and a gating feedforward network, and is used for carrying out context modeling and global dependency enhancement on the fusion features to obtain feature representations finally used for target identification and classification.
5. The method for identifying a radar HRRP deformation target based on feature correction of claim 4 wherein in S4, the parameters of the feature correction network are optimized by joint training with the paired sample sets as input data, including; s41, respectively converting the HRRP deformation target data in the paired sample sets With HRRP raw target data Extracting features through the feature extraction network to correspondingly obtain deformed feature representation And original characteristic representation ; S42, respectively And (3) with Correcting the space dimension through a space attention network to generate position level importance weights in the distance dimension, and correspondingly obtaining the characteristics after the space correction And (3) with Wherein the spatial attention network comprises two layers of one-dimensional convolution, a nonlinear activation function and a normalization function; S43, respectively And (3) with Correcting channel dimension through a channel attention network to model importance of each characteristic channel in the channel dimension, and correspondingly obtaining characteristics after channel correction And (3) with ; S44, sequentially and parallelly integrating two paths of characteristics And (3) with Self-adaptive weighted fusion is carried out through a three-branch attention network, and deformation HRRP characteristics after fusion are correspondingly obtained With original HRRP characteristics ; S45, respectively characterizing the deformation HRRP With original HRRP characteristics Feature denoising is carried out on the basis of an improved Pureformer network, so that feature representation finally used for target identification and classification is obtained; s46, carrying out forward propagation on the characteristic correcting network through the steps from S41 to S45, and calculating a composite loss function comprising various losses in the forward propagation process so as to iteratively optimize the parameters of the characteristic correcting network until training is completed.
6. The feature correction-based radar HRRP deformation target identification method of claim 5 wherein S42 includes: 42a) Will respectively And (3) with In an input space attention network, compressing the channel number of an input feature through a first layer of one-dimensional convolution, and extracting local space structure information under the condition of keeping the distance dimension unchanged to obtain a first space feature; 42b) Processing the first spatial feature by adopting a ReLU nonlinear activation function to obtain a second spatial feature; 42c) Mapping the second spatial feature into a single-channel spatial response through a second layer of one-dimensional convolution to obtain a third spatial feature; 42d) Normalizing the spatial response of the third spatial feature to a 0-1 interval by using a Sigmoid function, and correspondingly obtaining the spatially corrected feature And (3) with 。
7. The feature correction-based radar HRRP deformation target identification method of claim 5 wherein S43 includes: 43a) Will respectively And (3) with In an input channel attention network, carrying out global convergence on the characteristics of each channel through self-adaptive average pooling to obtain first channel characteristics; 43b) Compressing the channel dimension of the first channel characteristic by adopting one-dimensional convolution, and combining with a ReLU activation function to obtain a second channel characteristic; 43c) Mapping the second channel characteristics back to the original channel number through a second layer of one-dimensional convolution, generating response weights of all channels, and obtaining third channel characteristics; 43d) Normalizing the weight of the third channel characteristic to a 0-1 interval by using a Sigmoid function, and correspondingly obtaining the corrected characteristic of the channel And (3) with 。
8. The feature correction-based radar HRRP deformation target identification method of claim 5 wherein S44 includes: 44a) Features of two paths And (3) with Respectively inputting the two first fusion characteristics into a multi-scale convolution attention network, and carrying out fusion modeling comprising global average pooling operation and local one-dimensional convolution operation; 44b) The two first fusion features are subjected to channel compression convolution, batch normalization and ReLU activation treatment respectively, and attention response consistent with the number of original channels is generated through channel recovery convolution, so that two second fusion features are correspondingly obtained; 44c) Weighting and fusing the two second fusion features by using a Sigmoid function to correspondingly obtain fused deformation HRRP features With original HRRP characteristics 。
9. The method for feature correction based radar HRRP deformation target identification of claim 5 wherein in S45, the improvement based Pureformer network includes a feature enhancement module, a multi-headed depth separable self-attention module, and a gated feed forward network, then S45 includes: 45a) Respectively, will deform HRRP characteristic With original HRRP characteristics Inputting the first denoising feature into the feature enhancement module, performing multi-scale modeling on the input features through a parallel one-dimensional convolution structure, and performing fusion compression on the features with different scales to obtain the first denoising feature; 45b) Inputting the first denoising feature into the multi-head depth separable self-attention module, explicitly modeling the correlation between different positions in the distance dimension by utilizing a multi-head self-attention mechanism, and performing depth separable convolution and feature normalization operation to obtain a second denoising feature; 45c) And inputting the second denoising characteristic into the gating feedforward network, and realizing nonlinear characteristic transformation and inter-channel interaction through channel expansion, depth separable convolution and a gating mechanism to obtain characteristic representation finally used for target identification and classification.
10. The feature correction-based radar HRRP deformation target identification method of claim 5 wherein in S46 the composite loss function includes at least two of identity consistency loss, triple loss, classification loss, center loss, or contrast loss.

Description

Radar HRRP deformation target identification method based on feature correction Technical Field The invention belongs to the technical field of radar target detection and identification, and particularly relates to a radar HRRP deformation target identification method based on feature correction. Background The radar high-resolution range profile (High Resolution Range Profile, HRRP) is used as a one-dimensional electromagnetic scattering feature, and can accurately depict the projection distribution and scale information of a target scattering center along the radar sight line direction. By virtue of the advantages of convenience in data acquisition, low dimensionality, mature processing algorithm and the like, HRRP has become a key data source in the field of radar automatic target recognition (Radar Automatic Target Recognition, RATR). Currently, the existing recognition model is mostly built on an ideal detection hypothesis, namely, a preset high signal-to-noise ratio environment, a complete and non-shielding target structure, and a training set and a testing set are distributed consistently. However, in an actual non-cooperative target detection task, the ideal condition is often difficult to meet. Due to the complexity of battlefield environment and the variability of the target, distortion or missing of scattering information of a part of distance units often occurs in measured data, so that serious characteristic mismatch is generated between a test sample and a preset template. The deformation of the target mainly results from two factors, namely non-rigid change of the mounting configuration of the target, such as adjustment of weapon mounting or external components according to task requirements, direct change of electromagnetic scattering topology of the target, and information interception of a local structure of the target in the radar sight direction due to environmental shielding or camouflage interference. Particularly in a low signal-to-noise ratio environment, the additive interference of noise and the structural deformation of a target present complex coupling effects, so that effective scattering characteristics are further submerged, and the robustness of the traditional identification method faces serious challenges. Aiming at the problems, how to realize the accurate interpretation of HRRP under the double constraint of low signal-to-noise ratio and deformation interference becomes a technical difficulty to be overcome in the current radar signal processing field. In recent years, HRRP recognition research on deformation targets is getting more attention from the academic circles, and the existing representative methods mainly comprise the following two types, namely a variant target resolution range profile recognition method based on block sparse bayesian. For example, patent document 201810978483.0 discloses a 'variant target resolution range profile identification method based on block sparse Bayes', which constructs a variant target mathematical model, and solves and separates variant components in echo by using a block sparse Bayesian iterative algorithm by defining a variable prior probability and parameter distribution thereof. And then, the system eliminates the variant component, and carries out classification identification on the reconstructed pure range profile by using an adaptive Gaussian classifier. Although this approach weakens the interference of variant factors on recognition accuracy to some extent, it is still essentially a "recover before recognize" step-by-step processing architecture. This cascading process flow results in a final recognition performance that is highly dependent on the restoration quality of the front end, which, once there is a residual in the restoration phase, will directly result in distortion of the subsequent classification features. In addition, the end-to-end global optimization cannot be realized by the staged independent optimization, so that accumulation and transmission of system errors are easily caused, and further improvement of the overall recognition rate is limited. Another type is a method for identifying variant aircraft high-resolution range profile based on convolutional neural network. For example, patent document 201910201317.4 discloses a "convolutional neural network-based method for identifying variant aircraft high-resolution range profile", which aims to mine deep common features between variant and non-variant target HRRP data through a convolutional neural network, and uses the invariant features to alleviate feature mismatch problems caused by shape changes, so as to improve the robustness of identification. However, the algorithm has obvious limitations in practical application, on one hand, the characteristic extraction process is highly dependent on the integrity of target information, once a key scattering center area is shielded, the extraction of common characteristics is severely interfered to obviously