CN-122017837-A - ISAR small sample rapid imaging method and system based on stacked complex neural network

Abstract

The invention discloses an ISAR small sample rapid imaging method and system based on stacked complex neural networks, and belongs to the technical field of radar signal processing; the method comprises the steps of obtaining ISAR radar original echo data, carrying out multidimensional feature extraction to construct an enhanced feature vector, constructing a base model library by adopting a stacking integration strategy, utilizing K-fold cross validation to train a base model aiming at the real part and the imaginary part of an echo respectively, generating an out-of-bag predicted value as a meta-feature, constructing a full complex neural network as a meta-model, mapping physical coordinate features into high-frequency phase features through a Fourier feature mapping module, splicing the high-frequency phase features with the meta-feature, inputting the meta-model, constructing a mixed loss function, training the meta-model, retraining the base model, inputting data to be predicted into the trained base model and the meta-model, outputting a final complex echo predicted value, and carrying out ISAR imaging. The method ensures robustness by utilizing machine learning, and effectively solves the problems of phase fracture and high-frequency oscillation detail loss.

Inventors

• WAN TING
• RONG YUXIN

Assignees

• 南京邮电大学

Dates

Publication Date

20260512

Application Date

20260202

Claims (10)

1. 1. The ISAR small sample rapid imaging method based on stacked complex neural networks is characterized by comprising the following steps of: The method comprises the steps of obtaining ISAR radar original echo data, constructing a sample set, dividing the sample set into a training set and a testing set, extracting multidimensional features from the sample set, and constructing an enhanced feature vector containing physical coordinate features and local space statistical features; dividing a training set by using K-fold cross validation, respectively carrying out K-fold cross validation on the real part and the imaginary part of the echo to train a base model, generating an out-of-bag predicted value and taking the out-of-bag predicted value as a meta-feature; The physical coordinate features are mapped into high-frequency phase features through the Fourier feature mapping module, spliced with the meta features and input into the meta model; constructing a mixed loss function, wherein the mixed loss function comprises a data driving term used for restraining real and imaginary part numerical errors of a complex echo and a physical driving term used for restraining consistency of echo amplitude energy distribution; and retraining the base model by using the training set data, inputting the data of the test set into the trained base model and meta model, outputting a final complex echo predicted value, and performing ISAR imaging.
2. 2. The method for fast imaging of ISAR small samples based on stacked complex neural networks of claim 1, wherein the enhanced feature vector comprises: The physical coordinate features comprise frequency, azimuth angle, square of frequency, square of azimuth angle and interaction product term of frequency and azimuth angle; Local spatial statistics include local mean, local variance, gradient in the frequency direction, gradient in the azimuth direction, global mean, and phase variance.
3. 3. The method for rapidly imaging small ISAR samples based on stacked complex neural networks according to claim 1, wherein the base models in the base model library at least comprise a linear regression model, a support vector regression model and a random forest regression model.
4. 4. The ISAR small sample rapid imaging method based on stacked complex neural networks according to claim 1, wherein the fourier feature mapping module maps physical coordinate features to a formula of high frequency phase features as follows: Wherein, the For the normalized frequency and azimuth coordinate vector, To follow a gaussian distribution and untrainable random projection matrix, Is the output high-frequency phase eigenvector.
5. 5. The ISAR small sample rapid imaging method based on stacked complex neural networks of claim 1, wherein the network layer weights, biases and activation functions of the full complex neural networks are defined in complex domains, wherein complex linear layers define weights And input Its forward propagation operation follows complex multiplication rules: wherein the subscript And Representing real and imaginary parts, respectively, W is a weight matrix, x is an input feature vector, For bias, output is the linear transformation result.
6. 6. The method for fast imaging of small ISAR samples based on stacked complex neural networks of claim 1, wherein the mixing loss function is: Wherein, the Is the total mixing loss function value; For the balance coefficient, for adjusting the weight duty ratio of the physical driving term in the total loss; The method is a data driving term and is used for restraining numerical errors of real and imaginary parts of complex echoes; is a physical driving term used to constrain the envelope distribution of echo amplitude.
7. 7. An ISAR small sample rapid imaging system based on a stacked complex neural network, performing the method of any of claims 1-6, comprising: the data acquisition and feature extraction module is used for acquiring ISAR radar original echo data, constructing a sample set and dividing the sample set into a training set and a testing set; The meta-feature generation module is used for constructing a base model library containing a plurality of heterogeneous machine learning algorithms by adopting a stacking integration strategy, dividing a training set by utilizing K-fold cross validation, respectively carrying out K-fold cross validation on a real part and an imaginary part of an echo to train the base model, and generating an out-bag predicted value as a meta-feature; The feature fusion module is used for constructing a full complex neural network comprising a Fourier feature mapping module as a meta-model, mapping the physical coordinate features into high-frequency phase features through the Fourier feature mapping module, splicing the high-frequency phase features with the meta-features, and inputting the meta-model; The meta model training module is used for constructing a mixed loss function, comprising a data driving item for restraining real and imaginary part numerical errors of a complex echo and a physical driving item for restraining consistency of echo amplitude energy distribution, training the meta model in a training set by utilizing the mixed loss function, and updating meta model parameters through back propagation; And the prediction module retrains the base model by utilizing the training set data, inputs the data in the test set into the trained base model and meta model, outputs a final complex echo predicted value and performs ISAR imaging.
8. 8. A computer storage medium storing a readable program, wherein the program, when executed, is capable of instructing a computing device to perform the ISAR small sample rapid imaging method based on a stacked complex neural network as claimed in any one of claims 1 to 6.
9. 9. An electronic device is characterized by comprising a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface complete communication with each other through the communication bus; the memory is configured to store at least one executable instruction that causes the processor to perform operations corresponding to the stacked complex neural network-based ISAR small sample rapid imaging method according to any one of claims 1 to 6.
10. 10. A computer program product comprising computer instructions that instruct a computing device to perform operations corresponding to the stacked complex neural network based ISAR small sample rapid imaging method of any of claims 1-6.

Description

ISAR small sample rapid imaging method and system based on stacked complex neural network Technical Field The invention belongs to the technical field of radar signal processing, and particularly relates to an ISAR small sample rapid imaging method and system based on stacked complex neural networks. Background The inverse synthetic aperture radar (InverseSyntheticApertureRadar, ISAR) is an all-weather and all-day high-resolution microwave imaging technology, can acquire two-dimensional high-resolution images of non-cooperative moving targets, and has important application value in the fields of space monitoring, target identification, reverse conduction defense and the like. Conventional ISAR imaging algorithms, such as Range-Doppler (RD) and Back-Projection (BP) algorithms, are typically based on fast fourier transform or coherent accumulation principles. The method can obtain good imaging effect under the conditions of complete echo data and higher signal-to-noise ratio. However, in a practically complex electromagnetic environment, limited by the measurement time, storage capacity or hostile interference of the radar system, only echo data of a sparse aperture or a small sample is often obtained. In the case of incomplete data, the conventional imaging algorithm may have serious side lobe interference, grating lobe effect or resolution degradation, and it is difficult to form a clearly focused target image. In order to solve the imaging problem under sparse data, in recent years, a data-driven based machine learning and deep learning method is introduced into the field of radar signal processing. However, the existing related art still has the following significant drawbacks in practical applications: 1) The limitations of conventional machine learning methods are that although conventional algorithms such as Support Vector Machines (SVMs) or Random Forests (RFs) are robust under small samples, they generally cannot directly process complex data. The existing practice is often to train the real and imaginary parts of the radar echo as two independent real scalar quantities. The processing mode breaks the inherent phase coupling relation of the complex signals, so that the predicted result is distorted in phase, and the azimuth focusing performance of ISAR imaging is affected; 2) The problem of spectrum deviation of a deep neural network is that the deep learning method based on a multi-layer perceptron (MLP) or a Convolutional Neural Network (CNN) has strong nonlinear fitting capability, but a phenomenon of spectrum deviation (SpectralBias) is common, namely the neural network tends to learn low-frequency smooth components preferentially, and high-frequency oscillation details are difficult to fit. The radar echo signal is a typical high-frequency oscillation signal in nature, the high-frequency texture is often lost as a result of common network prediction, the imaging result is fuzzy, a group of detail-free 'light spots' are presented in a target area, a scattering center cannot be distinguished, and a deep learning model is usually trained by depending on massive data. In ISAR imaging tasks, the cost of acquiring a large number of marked electromagnetic simulation data or measured data is extremely high. Under the condition that only a small number of samples (such as 20% sparse rate) exist, a complex deep neural network is extremely easy to be subjected to overfitting, so that generalization capability is poor, and accurate prediction of echo with no angle can not be performed. Disclosure of Invention Aiming at the defects of the prior art, the invention aims to provide an ISAR small sample rapid imaging method and system based on stacked complex neural networks, which solve the problems in the prior art. The aim of the invention can be achieved by the following technical scheme: An ISAR small sample rapid imaging method based on stacked complex neural networks comprises the following steps: The method comprises the steps of obtaining ISAR radar original echo data, constructing a sample set, dividing the sample set into a training set and a testing set, extracting multidimensional features from the sample set, and constructing an enhanced feature vector containing physical coordinate features and local space statistical features; dividing a training set by using K-fold cross validation, respectively carrying out K-fold cross validation on the real part and the imaginary part of the echo to train a base model, generating an out-of-bag predicted value and taking the out-of-bag predicted value as a meta-feature; The physical coordinate features are mapped into high-frequency phase features through the Fourier feature mapping module, spliced with the meta features and input into the meta model; constructing a mixed loss function, wherein the mixed loss function comprises a data driving term used for restraining real and imaginary part numerical errors of a complex echo and a physical driving term used f