A method for generating joint anti-jamming radar signal processing of adversarial and subspace projection

Abstract

This invention discloses a joint anti-jamming radar signal processing method combining generative adversarial techniques (GAP) and subspace projection, relating to the fields of radar signal processing and anti-jamming technology. First, this invention decouples the transmitted waveform through frequency domain transformation, transforming the interference suppression problem into a blind source separation task. Then, it employs an autoencoder based on the U-net architecture for signal reconstruction and introduces a subspace projection mechanism in the latent space, explicitly eliminating interference components by utilizing the orthogonality of the target and interference on the latent manifold. Simultaneously, a composite loss function incorporating physical structure constraints and statistical latent space regularization is designed to ensure high fidelity of the reconstructed signal. Compared with existing technologies, this invention exhibits higher suppression efficiency and robustness in scenarios with low SINR, multiple interference types, and coexistence of clutter, providing a new technical approach for radar anti-jamming.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing and anti-jamming technology, and in particular to a combined anti-jamming radar signal processing method of generative adversarial techniques and subspace projection. Background Technology

[0002] Radar systems are primary sensors for environmental perception and target acquisition. However, the rapid development of Digital Radio Frequency Memory (DRFM) technology has brought unprecedented threats to radar systems. DRFM jammers can intercept and modulate radar transmitted signals to generate high-fidelity deception signals that are highly correlated with real targets. These signals maintain intra-pulse coherence, making it difficult for traditional matched filtering methods to distinguish between real and false targets. In particular, main lobe deception jamming, because the jamming signal and the target signal originate from the same spatial direction, renders traditional spatial filtering techniques completely ineffective, posing a severe challenge to target detection and tracking in radar systems.

[0003] Multiple-input multiple-output (MIMO) radar technology enhances system performance by utilizing multiple transmit and receive antennas, offering advantages in anti-jamming capabilities. MIMO radar can gain additional degrees of freedom through waveform diversity and spatial diversity, improving the system's spatial resolution and anti-jamming ability. However, MIMO radar also faces significant challenges when dealing with deceptive jamming in the main lobe direction, as the jamming signal and the target signal cannot be separated in the spatial domain. Summary of the Invention

[0004] The purpose of this invention is to provide a joint anti-jamming radar signal processing method combining generative adversarial techniques and subspace projection. This method is a MIMO radar main lobe interference suppression scheme. This invention can achieve high-fidelity target signal reconstruction in environments with low signal-to-interference-plus-noise ratio (SINR), complex clutter, and composite interference, thereby improving the anti-jamming performance and real-time processing capability of the radar system.

[0005] The technical solution adopted in this invention is: a joint anti-jamming radar signal processing method combining generative adversarial techniques and subspace projection, which includes the following steps:

[0006] Step 1, Signal Preprocessing and Frequency Domain Conversion: This involves processing the received time-domain signal... Sampling and analog-to-digital conversion are performed, and then a Fast Fourier Transform is applied to the resulting discretized signal to obtain its frequency domain representation. And based on the transmitted waveform spectrum In the frequency domain Decoupling is performed to obtain the frequency domain sequence. ;

[0007] Step 2, Encoder Feature Extraction and Latent Space Mapping: This involves mapping the frequency domain sequence... The input is fed into a subspace projection denoising autoencoder (SPDAE) based on the U-net architecture. The encoder of the SPDAE extracts high-level abstract features of the signal layer by layer, and outputs a one-dimensional latent space vector. The SPDAE encoder's input data includes two channels: a real part and an imaginary part. The spatial dimension of the input data is determined by the number of frequency domain sampling points. Determined; latent space vector It represents a hybrid representation of the target signal and interference signal in the depth feature space;

[0008] Step 3, Suppression of latent subspace projection interference: Based on the projection image matrix set offline. For latent space vectors A projection operation is performed to suppress interference components, resulting in a projection vector. , that is, the purified latent space vector;

[0009] Step 4, Decoder signal reconstruction and output: Project the vector The decoder input to the SPDAE undergoes layer-by-layer upsampling and reconstruction, outputting an estimated clean target frequency domain response of the same dimension as the encoder's input data. ;

[0010] Step 5, Signal Recovery and Output: The frequency domain response output by the decoder... With the transmitted waveform spectrum The estimated frequency domain signal is recovered by coupling, and then subjected to inverse fast Fourier transform back to the time domain to obtain a clean target signal with suppressed main lobe interference.

[0011] Furthermore, the loss function used by SPDAE during model training includes a delay frequency domain structure loss term. and latent space vector probability regularization loss .

[0012] Furthermore, the delay frequency domain structure loss term Specifically: ,in For inverse Fourier transform, Represents the mathematical expectation, that is Indicates time delay The expected value when the distribution is uniform from 0 to 1.

[0013] Furthermore, probability regularization loss Specifically: ,in It is the identity matrix. This is the penalty coefficient.

[0014] Furthermore, the encoder of SPDAE is specifically set as follows:

[0015] First, the input data's channel count is expanded to C1 feature channels using an initial convolutional layer. Then, primary feature extraction is performed through two downsampling blocks. The extracted primary feature map maintains C1 channels and its spatial dimension. Subsequently, a first spatial compression is performed through a downsampling convolutional layer, halving the spatial dimension. Next, a third downsampling block expands the number of feature channels from C1 to C2. Then, three more downsampling blocks are used for deep feature extraction, maintaining the number of channels in the extracted deep feature map at C2 and preserving the spatial dimension. Finally, at the bottleneck layer, the number of feature channels is compressed from C2 to 1 using the sixth downsampling block, and the feature tensor is reshaped into a one-dimensional latent space vector. ;

[0016] Each downsampling block includes a residual network module and a self-attention mechanism to capture the local structure and global dependencies of the signal. C1 and C2 are preset values.

[0017] Furthermore, the SPDAE decoder is specifically configured as follows:

[0018] First, the one-dimensional projection vector Reconstruct it into a three-dimensional tensor, restoring the spatial dimension to 1. Then, the number of feature channels is expanded from 1 to C2 through the first upsampling block; subsequently, feature refinement is performed through three consecutive upsampling blocks, with the spatial dimension of the feature maps output by the three upsampling blocks being C2. The number of feature channels is C2. Skip connections are set between the fifth downsampling block of the encoder and the second upsampling block of the decoder, between the fourth downsampling block of the encoder and the third upsampling block of the decoder, and between the third downsampling block of the encoder and the fourth upsampling block of the decoder. These skip connections perform feature-level fusion of the feature maps of the encoder and decoder at the corresponding stages to enhance the recovery of detailed information. Subsequently, the feature map output by the fourth upsampling block is spatially expanded by an upsampling transposed convolutional layer, expanding the spatial dimension to [value missing]. This effectively restores the spatial resolution of the signal. Finally, two upsampling blocks compress the number of feature channels from C2 to C1, and then the final convolutional layer compresses the number of feature channels from C1 to 2 to output the frequency domain response. Two of the channels represent the frequency domain response. The real and imaginary parts of the signal are preserved, maintaining the amplitude and phase information of the signal; the upsampling block includes a residual network module and a self-attention mechanism.

[0019] Furthermore, projection matrix Specifically:

[0020] Multiple clean target signal samples are collected to form a signal sample set; the latent space vector of the signal sample set is obtained through an encoder. Calculate the covariance matrix of the signal sample set. ;

[0021] For covariance matrix Perform eigenvalue decomposition Before selection The eigenvectors corresponding to the largest eigenvalues ​​form the projection matrix: ,in Let be the dimension of the equivalent potential target subspace, and Less than or equal to the latent space vector Spatial dimension (i.e., potential spatial dimension, number of frequency domain sampling points) Half of it.

[0022] The technical solution provided by this invention brings at least the following beneficial effects:

[0023] This invention first decouples the transmitted waveform through frequency domain transformation, transforming the interference suppression problem into a blind source separation task. Then, it employs an autoencoder based on a U-net architecture for signal reconstruction and introduces a subspace projection mechanism in the latent space, explicitly eliminating interference components by utilizing the orthogonality of the target and interference on the latent manifold. Simultaneously, a composite loss function incorporating physical structure constraints and statistical latent space regularization is designed to ensure high fidelity of the reconstructed signal. Compared with existing technologies, this invention exhibits higher suppression efficiency and robustness in scenarios with low SINR, multiple interference types, and coexisting clutter, providing a new technical approach for radar anti-jamming. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the overall model of a MIMO radar and SPDAE-based denoising processing, showing the structure of a co-located MIMO radar system, the spatial relationship between the target and the DRFM jammer, and the SPDAE processing flow.

[0026] Figure 2 The network structure diagram of the latent subspace projection denoising autoencoder (SPDAE) proposed in the embodiments of the present invention includes an encoder-decoder architecture based on U-net, a latent spatial projection layer, and a composite loss function;

[0027] Figure 3This is a flowchart illustrating the overall algorithm of the method of the present invention;

[0028] Figure 4 This is a schematic diagram of the processing procedure of the method of the present invention;

[0029] Figure 5 The interference suppression performance of this method is shown in the graphs for different single interference types, including: (a) ISRJ interference; (b) DFTJ interference; (c) ISFJ interference; and (d) VGJ interference. A comparison of pulse compression results is also shown under the conditions of SNR=-20dB and INR=10dB.

[0030] Figure 6 The interference suppression performance of this method under the condition of coexistence of clutter and multiple interferences is shown in the following diagrams: (a) the scenario of coexistence of target, CSJ and clutter; (b) the scenario of coexistence of target, ISRJ and clutter.

[0031] Figure 7 The graph shows the output SINR performance of the method of the present invention under different input SINR conditions. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0033] Consider a co-located MIMO radar system, such as Figure 1 As shown, the system consists of M uniformly spaced transmitting elements and N receiving elements. The transmitted signal of the m-th transmitting element within one pulse can be expressed as:

[0034]

[0035] in, The baseband waveform of the m-th array element is... The pulse width. Let t be the carrier frequency and t be the sampling time. The rectangle function is defined as follows:

[0036]

[0037] set up Let represent the transmitted signal vector of the M transmitting array elements, i.e.:

[0038]

[0039] set up Let N represent the echo signal vector of the N receiving array elements, i.e.:

[0040]

[0041] Considering various active interferences from DRFM and clutter from the ground or sea surface, the echo signal can be expressed as:

[0042]

[0043] in, These are the target echo, interference echo, and clutter echo, respectively. This represents background noise, which can be Gaussian noise or non-Gaussian noise.

[0044] set up The impulse response of the target. This is the equivalent impulse response of the DRFM-modulated interference signal. Let be the impulse response of the i-th clutter block. Then, the components in equation (5) can be expressed as:

[0045]

[0046] in, For the angle of the i-th target, For the target quantity, This represents the number of clutter blocks. These represent the steering vectors of the transmission and reception spaces of the i-th target, respectively. In this embodiment, it is assumed that the target is accompanied by a DRFM jammer (i.e., main lobe jammer) at the same angle, so the number and angle of the jamming sources are approximately the same as those of the target.

[0047] For a uniform linear array (with an element spacing of half a wavelength), the steering vector can be written as:

[0048]

[0049] According to the conventional radar workflow, beamforming is performed on formula (5) to obtain:

[0050]

[0051] in, A weighted vector is formed for the receiving beam, and each element is used to characterize the weight of each receiving array element.

[0052] Assuming the focus is on the angular region of the first target. ,make Substituting formula (6) into formula (8), we get:

[0053]

[0054] in, That is, the echo signal obtained through reception and processing. The weighted transmitted waveform is equivalent to the synthesized waveform propagating towards the target or interference direction; This is the weighted multi-channel independent noise.

[0055] Generally speaking, spatial filtering techniques are ineffective in suppressing main lobe deception interference. Therefore, the echo signal given by equation (9) is still affected by deception interference components. The signal is contaminated by residual clutter components. Furthermore, the transmitted waveform exhibits a convolutional form, meaning that target information and waveform parameters are inseparably mixed in the time domain. To address this issue, this invention starts from equation (9) and effectively decouples the echo model from the transmitted waveform by converting the signal to the frequency domain, thereby endowing the framework with generalization capabilities. Subsequently, frequency domain features are extracted using a deep neural network, a low-dimensional manifold representation of the target signal is learned in the latent space, and effective denoising and interference suppression are achieved using a subspace projection mechanism.

[0056] In one embodiment, the joint anti-jamming radar signal processing method of generative adversarial capabilities and subspace projection provided by the present invention is based on a latent subspace projection denoising autoencoder (SPDAE). Its core is to achieve effective separation of the target and interference through steps such as frequency domain preprocessing, deep learning feature extraction, subspace projection denoising, signal reconstruction, and inverse transformation. Its specific implementation process includes:

[0057] Step 1, Frequency domain transformation and waveform decoupling:

[0058] Consider the correlation between the transmitted waveform and the corresponding impulse response in the time domain of the received echo signal through convolution. To separate the target, interference, and clutter components, a domain transformation is required to convert the echo signal to the frequency domain. At this point, the transmitted waveform spectrum and the impulse response spectrum are in the form of a product, i.e.:

[0059]

[0060] in, They are respectively , , , , Fourier transform. for The Fourier transform of the signal is still approximately equal to white noise because the Fourier transform is a linear transform.

[0061] Typically, the radar transmit waveform is known, so the dependence on the transmit waveform spectrum can be removed from the echo spectrum. Dividing both sides of equation (10) by ,get:

[0062]

[0063] Use simplified notation Formula (11) can be rewritten as:

[0064]

[0065] in, Indicates the equivalent echo signal. Represents the equivalent noise component. This represents the equivalent interference components (including deceptive interference and clutter). Let be the mixing coefficients of the i-th objective. Let K be the time delay for the i-th target, and K be the number of point targets.

[0066] The objective of this invention is to suppress interference using an encoder-decoder architecture, starting from formula (12). In formula (12), only the structural form of the target signal is known a priori, while its time delay parameters and complex coefficients are unknown; The intentional modulation forms (including deceptive interference and clutter) are often unknown and difficult to express in closed form; This represents complex Gaussian white noise. Therefore, the objective of this invention is to address the presence of potential spatial contamination (i.e., subject to...). Under the condition of disturbance, reconstruct a target signal with a known structure from a one-dimensional complex signal contaminated by noise and interference.

[0067] Step 2, Encoder Feature Extraction:

[0068] set up Indicates the delay as The point target system response vector, i.e.:

[0069]

[0070] in, The number of data points (i.e., the number of data points in the Fast Fourier Transform used in Formula (10)). Assume... ,in Indicates the target delay. This represents a uniform distribution over (0,1), which means that if only the time delay of the radar echo is considered, the probability of the target echo appearing at any position in the delay domain is equal.

[0071] Using an autoencoder model architecture, the encoding and decoding processes can be characterized as follows:

[0072]

[0073]

[0074] in, For encoding functions, For decoding function, The latent space encoding vector of the input signal. As a potential spatial dimension, The sequence reconstructed by the decoder.

[0075] This invention employs a fully convolutional autoencoder based on the U-net architecture, utilizing skip connections to effectively fuse low-level and high-level features for signal reconstruction. The network processes complex-valued frequency domain signals (Equation 12), formatting them into two-channel tensors representing the real and imaginary parts. The encoder path progressively compresses the input into a low-dimensional latent vector through a series of "downblocks" (containing ResNet and attention mechanisms) and convolutional downsampling. .

[0076] See Figure 2 The encoder path employs a progressive downsampling architecture, gradually extracting multi-scale features and compressing spatial dimensions through "downsampling blocks" and "downsampling layers." The input tensor dimension is... ,in The batch size is 2, where 2 represents two channels (real and imaginary parts), 1 represents a single spatial dimension (one-dimensional signal), and 1024 represents the number of frequency domain sampling points. The encoder first expands the number of input channels from 2 to 128 feature channels through an initial convolutional layer, and then performs primary feature extraction through two downsampling blocks (blue arrows), maintaining the feature dimension at 1024. Each downsampling block contains a residual network module and a self-attention mechanism to capture the local structure and global dependencies of the signal. Subsequently, a first spatial compression is performed through a downsampling convolutional layer (red arrow), halving the spatial dimension from 1024 to 512, resulting in an output dimension of... Next, the number of feature channels is expanded from 128 to 256 through a third downsampling block, resulting in... Then, deep feature extraction is performed through three consecutive downsampling blocks, while maintaining the feature dimension as 1. Finally, at the bottleneck layer, the number of feature channels is compressed from 256 to 1 using a sixth downsampling block, resulting in... The feature tensor (containing real and imaginary information) is then reshaped into a one-dimensional latent vector. That is, potential spatial dimension Through the above layer-by-layer processing, the encoder realizes the transformation from high-dimensional complex-valued frequency domain signals. The nonlinear mapping to a low-dimensional latent vector contains the core structural information of the target signal while suppressing some noise and interference components.

[0077] Step 3, Projection of latent subspace

[0078] While autoencoders can learn target features unsupervised, their latent representations become unstable and lead to reconstruction distortion when data is limited or perturbations such as low SNR are present. This invention ensures good behavior of the latent space by imposing statistical constraints during training and employs a subspace projection technique during inference to project contaminated signals onto a clean signal subspace, thereby significantly reducing distortion.

[0079] This process assumes a potential disturbance vector. With the target latent vector Orthogonal. Therefore, the projection It can suppress all components orthogonal to the clean target manifold, where The projection matrix represents the potential target subspace.

[0080] set up Let represent the covariance matrix of all possible target latent vectors, i.e.:

[0081]

[0082] Through eigenvalue decomposition ,in The characteristic matrix, , Let be the eigenvectors. The projection matrix is ​​defined as:

[0083]

[0084] in, Let be the dimension of the equivalent potential target subspace, and .

[0085] Step 4, Decoder signal reconstruction

[0086] The reconstructed target signal can be obtained by updating formula (15) in the following way:

[0087]

[0088] The decoder path is symmetrical to the encoder, using "up-blocks" (containing residual and attention mechanisms) and transposed convolutions (upsampling) to reconstruct a clean, full-resolution target signal. .

[0089] See Figure 2 The decoder receives the latent vectors after subspace projection purification. As input, it is gradually restored to a complex-valued frequency domain signal with the same dimension as the input through a symmetrical upsampling stage. First, the one-dimensional latent vector is reshaped into a three-dimensional tensor. To restore the spatial dimension, the number of feature channels is expanded from 1 to 256 using the first upsampling block (gray arrow), resulting in... The upsampling block contains a residual network module and a self-attention mechanism to initially reconstruct the coarse-grained structure of the signal. Then, three consecutive upsampling blocks are used for feature refinement, maintaining the feature dimension as 1. Simultaneously, the feature maps of the corresponding stages of the encoder are fused with the feature maps of the current layer of the decoder through skip connections in a U-shaped network, enhancing the ability to recover detailed information. Subsequently, spatial expansion is performed through upsampling transposed convolutional layers (green arrows), doubling the spatial dimension from 512 to 1024, resulting in... This effectively restores the spatial resolution of the signal. Finally, after two upsampling blocks, the number of feature channels is compressed from 256 to 128, resulting in... Then, the final convolutional layer compresses the number of feature channels from 128 to 2, yielding the final output. Two channels represent the real and imaginary parts of the reconstructed signal, respectively, preserving the signal's amplitude and phase information. Through the multi-scale feature fusion of the above layer-by-layer processing and skip connections, the decoder transforms the low-dimensional latent vector purified by subspace projection into a single vector. Reconstructed into a high-dimensional complex-valued frequency domain signal It achieves a nonlinear inverse mapping from the potential feature space to the original signal space. The reconstructed signal retains the core structural features of the target signal (such as time delay, amplitude, and phase) while effectively suppressing interference and noise components.

[0090] Step 4, Design of composite loss function

[0091] Based on the known structural characteristics of the target signal in formula (12), this invention introduces a delay frequency domain structural loss term. For each delay Model response to target The prediction must conform to the specified "specific structure", that is:

[0092]

[0093] in, To from the frequency domain to delay domain Transformation (such as inverse Fourier transform).

[0094] Furthermore, the covariance matrix of the latent vectors should have a good condition number, therefore let:

[0095]

[0096] in, It is the identity matrix. This is the penalty coefficient.

[0097] The total training loss of SPDAE is:

[0098]

[0099] Step 5, Signal Recovery and Output

[0100] The reconstructed frequency domain signal is converted back to the time domain to obtain the final target signal. This is then achieved through inverse Fourier transform. Convert to time domain form and recouple the transmitted waveform spectrum:

[0101]

[0102] Output The target signal after noise reduction and interference suppression can be directly used for radar detection and tracking.

[0103] Compared with traditional main lobe interference suppression methods, the method of this invention has the following advantages:

[0104] (1) Explicit latent space cleanup: Unlike traditional autoencoders, this invention introduces a latent subspace projection mechanism to explicitly project the contaminated latent vectors onto a pre-learned clean target subspace, effectively eliminating interference components orthogonal to the target manifold. This explicit cleanup mechanism significantly improves reconstruction fidelity and avoids residual noise problems;

[0105] (2) Frequency Domain Decoupling Design: This invention transforms the convolutional mixing model into a blind source separation problem by converting the signal to the frequency domain and dividing it by the known transmitted waveform spectrum, effectively decoupling the transmitted waveform from the target / interference response. This design gives the framework good generalization ability, does not depend on a specific transmitted waveform form, and avoids the difficulty of the high correlation between the target and interference waveforms in time domain separation;

[0106] (3) Combination of physical constraints and statistical regularization: The composite loss function designed in this invention simultaneously includes physical structural constraints ( ) and statistical latent space regularization ( This ensures that the reconstructed signal not only fits the data but also follows the inherent physical characteristics of the radar echo, guaranteeing the robustness of the algorithm even under low SNR conditions.

[0107] (4) High-efficiency real-time processing: The present invention adopts a feedforward network structure, and the inference time is only 0.16 seconds (signal length L=1024), which meets the requirements of real-time radar signal processing.

[0108] (5) Multi-scenario adaptability: The present invention exhibits good performance in various interference types (DFJ, ISFJ, ISRJ, CSJ, RGJ, VGJ), clutter coexistence, composite interference and multi-target scenarios, and has strong generalization ability and practical value.

[0109] In one embodiment, using Figure 3 The overall algorithm flow shown was simulated. The input signal used was the MIMO radar single-channel time-domain echo signal after receiving beamforming processing; the core processing unit of the interference suppression system was a pre-trained latent subspace projection denoising autoencoder (SPDAE) network, whose network architecture... Figure 2 As shown in Table 1, the simulation experiments were conducted on a PC equipped with an i7-7700 CPU, a GTX1060 GPU, and 32GB of RAM. To comprehensively evaluate the model's performance, various simulation scenarios were used, and the specific parameters of each scenario are shown in Table 1.

[0110] Table 1 Target and Interference Signal Parameters for Training and Test Sets

[0111]

[0112] The dataset used for training and validating SPDAE was constructed synthetically. The total training dataset contains 8000 different time-domain radar echo waveforms. The network input was generated by randomly sampling and mixing various types of interference signals with the corresponding target echo signals to simulate realistic main lobe deception scenarios. The ground truth output of the network was defined as the target echo signal free of interference and noise components. The model was trained for 100 epochs under this configuration to ensure convergence and robust feature learning.

[0113] Based on the simulation scenarios shown in Table 1, see Figure 4 The specific processing steps of the method proposed in this invention include:

[0114] Step S1: Signal preprocessing and frequency domain conversion;

[0115] First, the received time-domain signal Sampling and analog-to-digital conversion are performed. Then, a Fast Fourier Transform (FFT) is applied to the resulting discretized signal to obtain its frequency domain representation. Using the known transmission waveform spectrum of the radar system. The echo is decoupled in the frequency domain to obtain a frequency domain sequence. ;in, Indicates Fast Fourier Transform;

[0116] Step S2: Encoder Feature Extraction and Latent Space Mapping

[0117] frequency domain sequence The encoder input to the SPDAE network (its structure is as follows) Figure 2 As shown on the left, high-level abstract features of the signal are extracted layer by layer, and finally the high-dimensional input data is compressed into a low-dimensional real-valued latent space vector. ,Right now Latent space vector It is a hybrid representation of the target signal and the interference signal in the depth feature space;

[0118] Step 3: Suppression of latent subspace projection interference

[0119] Before decoding, the latent space vector A projection operation is performed to suppress interference components. Projection matrix. Obtained offline: During the model training phase, a large number of clean target signal samples are used to obtain their latent vector set through the encoder, and the covariance matrix of the set is calculated according to formula (16). And perform eigenvalue decomposition. Before selection The eigenvectors corresponding to the largest eigenvalues ​​constitute the projection matrix. ;

[0120] During the model usage phase, the latent space vectors Projected onto this pure target subspace: To eliminate interference components in the latent vector that are orthogonal to the target signal subspace;

[0121] Step 4: Decoder signal reconstruction and output

[0122] The purified latent space vector The input is fed into the decoder of the SPDAE network to progressively upsample and reconstruct the low-dimensional latent vector, restoring it to the estimated frequency domain response of the clean target: ;

[0123] Step 5: Signal Recovery and Output

[0124] The frequency domain response output by the decoder With the transmitted waveform spectrum The coupling is re-established to recover the frequency domain signal containing waveform information. Finally, an inverse fast Fourier transform (IFFT) is performed on the signal to transform it back to the time domain, yielding the clean target signal with main lobe interference suppressed. It can be directly used in subsequent processing steps such as pulse compression, target detection, and tracking.

[0125] The simulation verification in this example shows that:

[0126] (1) Computational efficiency analysis: Under the condition of signal length L=1024, the single inference time of the method of the present invention is 0.16 seconds, indicating that the integration of the latent subspace projection mechanism does not bring significant additional computational burden compared with the standard autoencoder architecture. The method of the present invention achieves a good balance between performance and computational efficiency, meeting the needs of real-time radar signal processing.

[0127] (2) Single-point target single interference scenario: Figure 5 The performance evaluation of the method of this invention under various single interference scenarios is demonstrated. The experiments were conducted under challenging low SNR conditions, specifically SNR = -20dB (-18dB in the ISFJ scenario) and high INR = 10dB.

[0128] (3) Simulation under clutter and multiple interference conditions: Figure 6 The performance was demonstrated in clutter and complex interference scenarios. Figure 6 (a) shows the results in a scenario where the target, CSJ, and clutter coexist. The method of this invention can eliminate most of the clutter and interference energy, making the target peak identifiable; Figure 6 (b) shows the result of the target being masked by ISRJ and clutter. In scenarios with relatively low clutter energy, the method of the present invention effectively suppresses residual clutter and interference energy, and the quality of the recovered target waveform is significantly improved.

[0129] (5) SINR Improvement Analysis: Figure 7The performance curves of the output SINR of the method of the present invention as a function of the input SINR are shown. Throughout the entire test input SINR range (approximately -28 dB to 0 dB), the method of the present invention maintains a good output SINR. Even under severe interference conditions with an input SINR as low as -25 dB, the method of the present invention still maintains an output SINR above 6 dB, demonstrating robustness and stable signal recovery capability under low SINR conditions.

[0130] This invention is particularly suitable for high-fidelity target signal reconstruction in environments with low signal-to-interference-plus-noise ratio (SINR), complex clutter, and combined interference. It effectively addresses various main-lobe deception interference types, including dense false target jitter (DFJ), intermittent sample-and-forward jitter (ISFJ), intermittent sample-and-repeat jitter (ISRJ), comb spectrum jitter (CSJ), range gate drag jitter (RGJ), and velocity gate drag jitter (VGJ). This invention features moderate computational complexity and high processing efficiency, maintaining good target detection performance in complex electromagnetic environments with strong interference and clutter. It has significant application value in radar anti-jamming and radar target detection.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0132] The above descriptions are merely some embodiments of the present invention. For those skilled in the art, various modifications and improvements can be made without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for jointly processing anti-jamming radar signals using generative adversarial techniques and subspace projection, characterized in that, Includes the following steps: Step 1, Signal Preprocessing and Frequency Domain Conversion: This involves processing the received time-domain signal... Sampling and analog-to-digital conversion are performed, and then a Fast Fourier Transform is applied to the resulting discretized signal to obtain its frequency domain representation. And based on the transmitted waveform spectrum In the frequency domain Decoupling is performed to obtain the frequency domain sequence. ; Step 2, Encoder Feature Extraction and Latent Space Mapping: This involves mapping the frequency domain sequence... The input is fed into a subspace projection denoising autoencoder (SPDAE) based on the U-net architecture. The encoder of the SPDAE extracts high-level abstract features of the signal layer by layer, and outputs a one-dimensional latent space vector. The SPDAE encoder's input data includes two channels: a real part and an imaginary part. The spatial dimension of the input data is determined by the number of frequency domain sampling points. Determined; latent space vector It represents a hybrid representation of the target signal and interference signal in the depth feature space; Step 3, Suppression of latent subspace projection interference: Based on the projection image matrix set offline. For latent space vectors A projection operation is performed to suppress interference components, resulting in a projection vector. , that is, the purified latent space vector; Step 4, Decoder Signal Reconstruction and Output: The projection vector... The decoder input to the SPDAE undergoes layer-by-layer upsampling and reconstruction, outputting an estimated clean target frequency domain response of the same dimension as the encoder's input data. ; Step 5, Signal Recovery and Output: The frequency domain response output by the decoder... With the transmitted waveform spectrum The estimated frequency domain signal is recovered by coupling, and then subjected to inverse fast Fourier transform back to the time domain to obtain a clean target signal with suppressed main lobe interference.

2. The method as described in claim 1, characterized in that, The loss function used by SPDAE during model training includes a delay-frequency domain structure loss term. and latent space vector probability regularization loss .

3. The method as described in claim 2, characterized in that, Delay frequency domain structure loss term Specifically: ,in For inverse Fourier transform, Indicates time delay The expected value when the distribution is uniform from 0 to 1.

4. The method as described in claim 2, characterized in that, Probabilistic regularization loss Specifically: ,in It is the identity matrix. This is the penalty coefficient.

5. The method as described in claim 1, characterized in that, The encoder settings for SPDAE are as follows: First, the input data's channel count is expanded to C1 feature channels using an initial convolutional layer. Then, primary feature extraction is performed through two downsampling blocks. The extracted primary feature map maintains C1 channels and its spatial dimension. Subsequently, a first spatial compression is performed through a downsampling convolutional layer, halving the spatial dimension. Next, a third downsampling block expands the number of feature channels from C1 to C2. Then, three more downsampling blocks are used for deep feature extraction, maintaining the number of channels in the extracted deep feature map at C2 and preserving the spatial dimension. Finally, at the bottleneck layer, the number of feature channels is compressed from C2 to 1 using the sixth downsampling block, and the feature tensor is reshaped into a one-dimensional latent space vector. ; Each downsampling block includes a residual network module and a self-attention mechanism, with C1 and C2 being preset values.

6. The method as described in claim 5, characterized in that, The SPDAE decoder is specifically configured as follows: First, the one-dimensional projection vector Reconstruct it into a three-dimensional tensor, restoring the spatial dimension to 1. Then, the number of feature channels is expanded from 1 to C2 through the first upsampling block; subsequently, feature refinement is performed through three consecutive upsampling blocks, with the spatial dimension of the feature maps output by the three upsampling blocks being C2. The number of feature channels is C2. Skip connections are set between the fifth downsampling block of the encoder and the second upsampling block of the decoder, between the fourth downsampling block of the encoder and the third upsampling block of the decoder, and between the third downsampling block of the encoder and the fourth upsampling block of the decoder. These skip connections are used to fuse the feature maps of the encoder and decoder at the corresponding stages. Then, the feature map output by the fourth upsampling block is spatially expanded by an upsampling transposed convolutional layer, expanding the spatial dimension to [value missing]. Finally, the number of feature channels is compressed from C2 to C1 through two upsampling blocks, and then compressed from C1 to 2 through the final convolutional layer to output the frequency domain response. Two of the channels represent the frequency domain response. The real and imaginary parts of the sampled data are included in the upsampling block, which consists of a residual network module and a self-attention mechanism.

7. The method according to any one of claims 1 to 6, characterized in that, Projection matrix Specifically: Multiple clean target signal samples are collected to form a signal sample set; the latent space vector of the signal sample set is obtained through an encoder. Calculate the covariance matrix of the signal sample set. ; For covariance matrix Perform eigenvalue decomposition Before selection The eigenvectors corresponding to the largest eigenvalues ​​form the projection matrix: ,in Let be the dimension of the equivalent potential target subspace, and Less than or equal to the latent space vector Spatial dimension.

Images