Radar adaptive jamming cancellation method based on deep time-frequency network

Abstract

The present disclosure relates to the field of radar electronic countermeasures, and provides a radar adaptive jamming cancellation method based on a deep time-frequency network, comprising: obtaining a mixed signal obtained by a main radar receiving end as a main signal, and obtaining a reference signal obtained by an auxiliary radar receiving end as an auxiliary signal, and performing signal alignment on the main signal and the auxiliary signal; performing short-time Fourier transform on the main signal and the aligned auxiliary signal to obtain a main signal complex time-frequency spectrum and an auxiliary signal complex time-frequency spectrum, decomposing both into two components of real and imaginary parts, and splicing the two components in the channel dimension to obtain a four-channel tensor; inputting the four-channel tensor into a deep neural network based on a symmetric encoder-decoder architecture and integrated with a self-attention mechanism to obtain a target signal complex time-frequency spectrum; reconstructing the target signal complex time-frequency spectrum back to a time-domain signal to obtain a target echo estimation signal after suppressing the homologous interference. The present disclosure can achieve more accurate interference suppression and target reservation in a complex interference environment.

Description

Technical Field

[0001] This disclosure relates to the field of radar electronic countermeasures technology, and in particular to a radar adaptive interference cancellation method based on deep time-frequency networks. Background Technology

[0002] Radar systems detect, locate, and identify targets by emitting electromagnetic waves and receiving their echoes, playing an irreplaceable role in modern military, meteorological, and aviation fields. However, in practical applications, radar receivers, in addition to receiving the desired target echoes, are also subject to interference from the environment or hostile jammers. This interference severely restricts radar performance. Especially in complex electromagnetic environments, this interference can severely obscure small targets, leading to a sharp decline in radar detection performance and even system failure.

[0003] In complex electromagnetic environments, interference suppression techniques can effectively suppress interference and improve the anti-interference performance of radar systems. Among various interference suppression techniques, adaptive interference cancellation technology has become a research hotspot due to its good real-time performance and adaptability. The core idea of ​​this technology is to use a reference channel to collect interference samples, and then adjust the weights through an adaptive filter to match the output with the interference signal in the main channel, thereby achieving interference cancellation. Traditional adaptive interference cancellation technology has been fully developed. For example, Hao et al. proposed an interference cancellation method based on a tracking algorithm, which significantly improved the isolation in static environments by utilizing the sparsity of the interference coupling path. For dynamic systems, Kalman filtering (KF) and its improved algorithms are widely used. Hao et al. further proposed the Sparse Constraint KF algorithm (SCKF), which improves the identification accuracy of dynamic sparse systems by introducing sparsity regularization, increasing the isolation by 3dB–5dB. In scenarios with high real-time requirements, Cao Lichang et al. proposed a fast adaptive interference cancellation algorithm for aerospace telemetry and control radar. By optimizing the convolution operation of the Least Mean Squares (LMS) algorithm, they reduced the computational load and kept the signal-to-noise ratio degradation within 0.3 dB. Liu Yemin et al. proposed an interference cancellation method based on multi-channel synthetic aperture radar (SAR). Through phase compensation and channel difference processing, they effectively suppressed the composite interference of multiple jammers. However, its performance depends on prior knowledge of the jammer positions and suffers from target loss period (TLP) phenomenon. Zhao Zhongkai et al. studied an adaptive cancellation technique based on the Variable Step-Size Affine Projection (VSSAP) algorithm for co-channel interference problems. By improving the step size function, they accelerated the convergence speed while reducing steady-state error. While traditional adaptive cancellation techniques have made significant progress, their performance is limited by the filter order, convergence algorithm, and assumptions about the statistical characteristics of the interference signal. For example, the LMS algorithm, though simple to implement, has slow convergence speed and large steady-state error; the Affine Projection (AP) algorithm improves convergence but is computationally complex; and sparse methods, while improving accuracy, are sensitive to the assumption of environmental sparsity, resulting in performance degradation in non-sparse scenarios. Furthermore, these methods often rely on linear models, making them difficult to handle non-stationary interference and complex nonlinear coupling problems. Traditional adaptive filtering methods, such as the LMS and Normalized Least Mean Squares (NLMS) algorithms, heavily rely on the assumption of stationarity of the interference signal and linear models. In real-world complex electromagnetic environments, interference signals often exhibit strong non-stationary and nonlinear characteristics, leading to slow convergence speed, poor tracking performance, or even failure of traditional algorithms, making accurate cancellation difficult.

[0004] With the improvement of digital signal processing capabilities, time-frequency analysis techniques such as Short-Time Fourier Transform (STFT) and wavelet transform have been introduced to better characterize the time-varying characteristics of interference. This enables interference identification algorithms to more accurately capture dynamic interference patterns such as single-frequency interference, multi-frequency interference, and linear frequency sweeping interference, laying the foundation for subsequent deep learning processing. Meanwhile, the success of machine learning techniques, especially deep learning, in image recognition and speech separation has made its application in radar signal processing possible. Zhao Jizhe et al. used complex neural networks to directly process the received signal in complex form, while simultaneously modeling amplitude and phase characteristics, achieving a signal-to-interference-plus-noise ratio (SNR) gain of over 56 dB under strong interference conditions. Zhang Wenchao et al. introduced the Convolutional Block Attention Module (CBAM) and combined it with the lightweight MobileNet to form CBAM-M-Net. By enhancing feature extraction capabilities through channel and spatial attention mechanisms, the interference recognition accuracy was improved to 99.86%, effectively overcoming the limitations of traditional Convolutional Neural Networks (CNNs) in complex channel environments. Furthermore, they adopted the U-Net architecture to reconstruct interference signals, thereby suppressing odd-frequency interference.

[0005] These advancements demonstrate that deep learning methods are evolving from general image processing to specialized signal processing. By combining domain knowledge (such as complex signal processing and time-frequency analysis) with novel network modules (such as attention mechanisms), they are gradually addressing the bottleneck issues of traditional CNNs in interference suppression, providing more robust solutions for radar and communication systems in highly dynamic environments. However, despite existing research attempting to apply deep learning to signal processing, most methods suffer from simple network structures and limited feature extraction capabilities. They often directly process time-domain signals, failing to fully utilize the structured features of interference in the time-frequency domain, and lack mechanisms for modeling global dependencies, resulting in insufficient generalization ability and accuracy under strong interference and complex scenarios. Summary of the Invention

[0006] This disclosure aims to address at least one of the problems existing in the prior art by providing a radar adaptive interference cancellation method based on deep time-frequency networks.

[0007] This disclosure provides a radar adaptive interference cancellation method based on deep time-frequency networks, the radar adaptive interference cancellation method based on deep time-frequency networks includes: The mixed signal obtained from the main radar receiver is used as the main signal, and the reference signal obtained from the auxiliary radar receiver is used as the auxiliary signal. The main signal and the auxiliary signal are then aligned to obtain the aligned auxiliary signal. The main signal and the aligned auxiliary signal are subjected to short-time Fourier transform processing respectively to obtain the complex time spectrum of the main signal and the complex time spectrum of the auxiliary signal. The complex time spectrum of the main signal and the complex time spectrum of the auxiliary signal are decomposed into two components, real part and imaginary part, and then concatenated in the channel dimension to construct a four-channel tensor. The four-channel tensor input is based on a deep neural network with a symmetric encoder-decoder architecture and an integrated self-attention mechanism to obtain the complex time spectrum of the target signal output by the deep neural network. The complex time spectrum of the target signal is reconstructed back into the time domain signal through short-time inverse Fourier transform, and the time domain signal is used as the target echo estimation signal after suppressing co-source interference.

[0008] Optionally, the step of performing signal alignment processing on the main signal and the auxiliary signal to obtain the aligned auxiliary signal includes: The main signal and the auxiliary signal are aligned using a generalized cross-correlation-phase transform weighting algorithm.

[0009] Optionally, the step of performing signal alignment processing on the main signal and the auxiliary signal using a generalized cross-correlation-phase transform weighting algorithm includes: The cross-power spectrum between the main signal and the auxiliary signal is determined according to the following formula. : ; in, The Fourier transform of the main signal, The Fourier transform of the auxiliary signal The complex conjugate; The PHAT weighting function is determined according to the following formula. : ; The weighted cross-power spectrum is determined according to the following formula. : ; According to the following formula, the weighted cross-power spectrum Performing the inverse Fourier transform yields the generalized cross-correlation function. : ; in, Indicates the inverse Fourier transform. This indicates the transmission delay between the main channel and the auxiliary channel; generalized cross-correlation function The corresponding value when the maximum value is obtained The value of As the estimated time delay, according to the following formula, using... Time delay compensation is performed on the auxiliary signal to obtain the aligned auxiliary signal. : ; in, Indicates the current moment. express Auxiliary signals for timing.

[0010] Optionally, the four-channel tensor is represented as: ; in, For the four-channel tensor and , The dimension is The real space, Indicates the height of the time-frequency image. Indicates the width of the time-frequency image. This indicates a splicing operation. The real component of the spectrum of the main signal in its complex form is represented. This represents the imaginary part of the spectrum of the main signal when it is complex. This represents the real component of the complex frequency spectrum of the auxiliary signal. This represents the imaginary component of the spectrum of the auxiliary signal when it is complex.

[0011] Optionally, the deep neural network includes an encoder and a decoder that is mirror-symmetric to the encoder, with an intermediate path provided between the encoder and the decoder.

[0012] Optionally, the encoder includes four cascaded downsampling blocks, each of which contains a ResNet layer, an activation function SiLU, and a downsampling operation layer. The downsampling process of the downsampling block is represented as follows: ; in, Indicates the first The output feature map of the layer, Indicates the first The output feature map of the layer, Indicates the first The downsampling block of the layer, The values ​​are 1, 2, 3, and 4 respectively. hour, It is a four-channel tensor; The latter two downsampling blocks in the encoder also each contain a multi-head self-attention layer, the calculation process of which includes: According to the following formula, for input features Perform a linear transformation to obtain the corresponding query matrix. Key matrix Sum matrix : ; in, The dimension is The real space, Indicates channel dimension, Represents the total number of spatial locations and , Indicates the height of the time-frequency image. Indicates the width of the time-frequency image. These are the query weight matrix, key weight matrix, and value weight matrix, respectively. , The dimension is The real space, Indicates the projection dimension; query matrix Key matrix Sum matrix Divided into Each head has a dimension of [number]. For the first For each size, calculate its attention output according to the following formula: ; in, Indicates the header number and , Indicates the first Attention output based on size express function, They represent the first The corresponding query matrix, key matrix, and value matrix for each head. Indicates matrix transpose; According to the following formula, the final self-attention output is obtained by concatenating the outputs of all heads and performing a linear transformation: ; in, This represents the final self-attention output. This indicates that the attention output of the first head will be... To the Attention output of size To splice, Indicates the output weight matrix and , The dimension is The real number space.

[0013] Optionally, the decoder includes four cascaded upsampling blocks, the upsampling process of which is represented as follows: ; in, Indicates the first Decoder features of the layer Indicates the first Decoder features of the layer Indicates the first The upsampling block of the layer, This indicates a channel splicing operation implemented by skip connections, when hour, The intermediate path is the output feature map of the 4th layer. The feature map output after processing; The two upsampling blocks in the decoder that are mirror images of the two downsampling blocks in the encoder, both of which contain the multi-head self-attention layer, also each contain the multi-head self-attention layer.

[0014] Optionally, the intermediate path includes a ResNet layer.

[0015] Optionally, the loss function of the deep neural network Represented as: ; in, Representing deep neural networks The parameters, Represents the number of training samples. Indicates the training sample number. Indicates the first The real target time-frequency tensor corresponding to each training sample Indicates the first The four-channel tensor corresponding to each training sample Indicates will Input deep neural network The obtained time-frequency tensor of the predicted target, This represents the L2 norm.

[0016] Optionally, reconstructing the complex time spectrum of the target signal back to the time domain signal via short-time inverse Fourier transform includes: According to the following formula, the real and imaginary parts of the complex time spectrum of the target signal are combined into a complex time-frequency matrix of the complex time spectrum of the target signal. : ; in, This represents the real part of the spectrum of the target signal when it is complex. This represents the imaginary part of the spectrum of the target signal when it is complex. Represents the imaginary unit; According to the following formula, the complex time spectrum of the target signal can be reconstructed back into the time domain signal through short-time inverse Fourier transform: ; in, The reconstructed result The time-domain signal at time t, express The short-time Fourier transform and , express The window function at time t, This indicates the transmission delay between the main channel and the auxiliary channel. Indicates frequency, Represents the natural constant.

[0017] The radar adaptive interference cancellation method based on deep time-frequency networks disclosed in this disclosure, compared with existing technologies, uses a four-channel tensor of the time-frequency signal as a conditional input and directly regresses the target output through an end-to-end deep neural network based on a symmetric encoder-decoder architecture and integrating a self-attention mechanism. It does not rely on traditional iterative algorithms or explicit linear filtering models, exhibiting strong nonlinear mapping and feature separation capabilities. This deep neural network can learn and fit the complex mixing relationship between interference and echo signals in the time-frequency domain, thereby achieving more accurate interference estimation and elimination, obtaining a higher Interference Cancellation Ratio (ICR) and lower signal distortion, significantly improving the radar's detection performance in strong interference environments. Furthermore, the forward inference process of the radar adaptive interference cancellation method based on deep time-frequency networks disclosed in this disclosure is calculated in one step using the optimized deep neural network, resulting in fast computation speed, low latency, and easy acceleration on existing graphics processing units (GPUs) or dedicated artificial intelligence (AI) chips, meeting the real-time processing requirements of radar systems. Attached Figure Description

[0018] One or more embodiments are illustrated by way of example with the corresponding pictures in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0019] Figure 1 A flowchart illustrating an adaptive radar interference cancellation method based on a deep time-frequency network, as provided in one embodiment of this disclosure. Figure 2 A schematic diagram of the core architecture of a deep neural network based on a symmetric encoder-decoder architecture and integrating a self-attention mechanism, provided for another embodiment of this disclosure. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the various embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this disclosure to facilitate a better understanding of the disclosure. However, the technical solutions claimed in this disclosure can be implemented even without these technical details and with various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of this disclosure. The various embodiments can be combined with and referenced by each other without contradiction.

[0021] In the context of existing technologies, the U-Net architecture, which combines an encoder-decoder structure with skip connections, has become an ideal candidate for solving the detail preservation problem in interference cancellation due to its ability to preserve details and integrate context in biomedical image segmentation. Furthermore, the self-attention mechanism in natural language processing, which can effectively model global dependencies in sequences, has been improved and applied to vision and signal processing tasks. This disclosure applies an optimized U-Net deep network model incorporating a self-attention mechanism to a radar interference cancellation task. Through an end-to-end time-frequency domain learning approach, this disclosure overcomes the linear assumptions and convergence bottlenecks of traditional methods, achieving more accurate interference suppression and target preservation in complex interference environments, opening up new research directions for radar anti-jamming technology.

[0022] One embodiment of this disclosure provides a radar adaptive interference cancellation method based on a deep time-frequency network. Its core lies in using a deep neural network model specifically designed for time-frequency domain signal processing to achieve high-precision separation and suppression of interference from mixed signals while preserving the target echo signal. The overall process of the radar adaptive interference cancellation method based on a deep time-frequency network provided in this embodiment is as follows: Figure 1 As shown, this includes steps one through four, which will be discussed below. Figure 1 Steps one through four are explained in detail.

[0023] Step 1: Interference signal alignment: Acquire the mixed signal obtained from the main radar receiver as the main signal, and at the same time acquire the reference signal obtained from the auxiliary radar receiver as the auxiliary signal. Perform signal alignment processing on the main signal and the auxiliary signal to obtain the aligned auxiliary signal.

[0024] Specifically, the main radar receiver and the auxiliary radar receiver receive signals simultaneously. The mixed signal obtained by the main radar receiver can be expressed as: ; in, Obtained from the main radar receiver The mixed signal at time, and also for The main signal at that moment; for The target echo signal at that moment, for Interference signal in the main channel at any given time. for Additive noise in the main channel at any given moment.

[0025] The reference signal obtained by the auxiliary radar receiver can be expressed as: ; in, For auxiliary radar receiver to obtain The reference signal at time, and also for Auxiliary signals for timing; for The main channel interference signal at any time, and also for Interference signals from the auxiliary channel at any given time; Transmission delay between the primary channel and the auxiliary channel, for Additive noise in the constant auxiliary channel.

[0026] For example, in step one, the main signal and the auxiliary signal are aligned to obtain the aligned auxiliary signal, including: using the Generalized Cross-Correlation with Phase Transform (GCC-PHAT) algorithm to align the main signal and the auxiliary signal.

[0027] Specifically, to eliminate the transmission delay between the main channel and the auxiliary channel To mitigate the impact of signal alignment, the GCC-PHAT algorithm is employed.

[0028] The generalized cross-correlation-phase transform weighting algorithm is used to perform signal alignment processing on the main signal and auxiliary signal, specifically including: The cross-power spectrum between the main signal and the auxiliary signal is determined according to the following formula. : ; in, Fourier transform of the main signal, Fourier transform of auxiliary signal The complex conjugate; The PHAT weighting function is determined according to the following formula. : ; The weighted cross-power spectrum is determined according to the following formula. : ; According to the following formula, the weighted cross-power spectrum Performing the inverse Fourier transform yields the generalized cross-correlation function. : ; in, Indicates the inverse Fourier transform. This indicates the transmission delay between the main channel and the auxiliary channel; generalized cross-correlation function The corresponding value when the maximum value is obtained The value of As the estimated time delay, according to the following formula, using... Time delay compensation is performed on the auxiliary signal to obtain the aligned auxiliary signal. : ; in, Indicates the current moment. express Auxiliary signals for timing.

[0029] Step 2: Time-frequency transformation: Perform short-time fourier transform (STFT) on the main signal and the aligned auxiliary signal respectively to map them from the one-dimensional time domain to the two-dimensional time-frequency domain, so as to reveal the non-stationary characteristics of the signal and obtain the complex time spectrum of the main signal and the complex time spectrum of the auxiliary signal.

[0030] For example, for Time signal Its short-time Fourier transform Defined as: ; in, for The signal of the moment, for Window functions for time intervals, such as the Hanning window, for The window function at time t, Indicates frequency, Represents the natural constant. It represents the imaginary unit.

[0031] Will and As respectively After performing a short-time Fourier transform, both the main signal and the auxiliary signal are represented as complex time-frequency matrices: ; ; in, The complex time-frequency spectrum matrix corresponding to the complex time-frequency spectrum of the main signal. Let be the complex time-frequency spectrum matrix of the target echo signal. The complex time-spectrum matrix of the main channel interference signal. The complex time-spectral matrix of the additive noise in the main channel. This is the complex time-frequency spectrum matrix corresponding to the complex time-frequency spectrum of the auxiliary signal. For the complex time-spectrum matrix of the interference signal in the auxiliary channel, The complex time-spectral matrix of the additive noise of the auxiliary channel.

[0032] Subsequently, the complex time spectrum of the main signal and the complex time spectrum of the auxiliary signal are both decomposed into real and imaginary components, and then spliced ​​together in the channel dimension to construct a four-channel tensor.

[0033] Specifically, the complex time-frequency spectrum of the main signal is decomposed into real components. and imaginary part quantity Simultaneously, the complex time spectrum of the auxiliary signal is decomposed into real components. and imaginary part quantity These four components are concatenated along the channel dimension to construct a four-channel deep neural network input tensor, i.e., a four-channel tensor. .

[0034] For example, a four-channel tensor is represented as: ; in, It is a four-channel tensor and , The dimension is The real space, Indicates the height of the time-frequency image. Indicates the width of the time-frequency image. This indicates a splicing operation. This represents the real part of the spectrum of the main signal when it is a complex number. This represents the imaginary part of the spectrum when the main signal is complex. This represents the real part of the spectrum of the auxiliary signal when it is complex. It represents the imaginary part of the spectrum when the auxiliary signal is complex.

[0035] By constructing the complex time-frequency spectrum of the main signal and the complex time-frequency spectrum of the auxiliary signal into a four-channel time-frequency image tensor, the radar interference cancellation problem is transformed into a pixel-level image-to-image regression problem in the time-frequency image domain. The goal of the deep neural network is to learn a mapping function from the four-channel input to the two-channel output. .

[0036] Step 3: Deep Neural Network: Input the four-channel tensor into a deep neural network based on a symmetric encoder-decoder architecture and integrating a self-attention mechanism to obtain the complex time spectrum of the target signal output by the deep neural network.

[0037] Specifically, the constructed four-channel tensor The input is a deep neural network based on a symmetric encoder-decoder architecture and integrating a self-attention mechanism. In the context of deep neural networks The parameters are Deep neural networks The goal is to estimate the time spectrum of the target signal as accurately as possible.

[0038] The deep neural network provided in this embodiment belongs to an improved U-Net architecture for time-frequency signals, supporting fine prediction of time-frequency images. For example, the deep neural network includes an encoder and a decoder that is mirror-symmetric to the encoder, with an intermediate path provided between the encoder and the decoder. The following describes... Figure 2 This deep neural network will be explained in detail.

[0039] For example, the encoder provides a downsampling path comprising four cascaded downsampling blocks, each containing a ResNet layer, a SiLU activation function, and a downsampling operation layer. Specifically, the ResNet layer is connected to the downsampling operation layer. The ResNet layer contains two 2D residual network blocks, i.e., ResNetBlock2D, with the output of the previous ResNetBlock2D layer serving as the input to the next. The downsampling operation layer is a 2D downsampling layer, i.e., DownSample2D, specifically a convolutional layer with a stride of 2. The input to the 2D downsampling layer is the output of the last 2D residual network block.

[0040] The downsampling process of the downsampling block is represented as follows: ; in, Indicates the first The output feature map of the layer, Indicates the first The output feature map of the layer, Indicates the first The downsampling block of the layer, The values ​​are 1, 2, 3, and 4 respectively. hour, It is a four-channel tensor. That is, the input of the downsampling block of the first layer is a four-channel tensor constructed using the complex time spectrum of the main signal and the complex time spectrum of the auxiliary signal. .

[0041] The last two downsampling blocks in the encoder, namely the third and fourth layers, are integrated self-attention 2D convolutional downsampling blocks, also known as attention downsampling blocks (AttnDownBlock). These attention downsampling blocks introduce a self-attention mechanism on top of convolutional feature extraction, including multi-head self-attention layers. They combine... Figure 2The attention downsampling block introduces a multi-head self-attention layer between the ResNet layer and the downsampling operation layer. The input of this multi-head self-attention layer is the output of the ResNet layer, which is also the output of the last ResNetBlock2D layer in the ResNet layer. The output of this multi-head self-attention layer is the input of the downsampling operation layer, i.e., DownSample2D.

[0042] The calculation process of the multi-head self-attention layer includes: First, satisfy Input features Remodeling to meet Input features ,in, The dimension is The real space, Indicates channel dimension, Indicates the height of the time-frequency image. Indicates the width of the time-frequency image. Represents the total number of spatial locations and , The dimension is The real number space, and then, according to the following formula, for the input features Perform a linear transformation to obtain the corresponding query matrix. Key matrix Sum matrix : ; in, These are the query weight matrix, key weight matrix, and value weight matrix, respectively. , The dimension is The real space, Indicates the projection dimension; query matrix Key matrix Sum matrix Divided into Each head has a dimension of [number]. ,get: , k = 1, …, h ; in, Indicates the header number and , They represent the first The query matrix and key matrix corresponding to each head. The dimension is The real number space; For the For each size, its self-attention weight matrix is ​​calculated according to the following formula: ; in, Indicates the first Self-attention weight matrix for size, express function, Indicates matrix transpose; For the For each size, calculate its attention output according to the following formula: ; in, Indicates the first Attention output based on size Indicates the first The value matrix corresponding to each head; According to the following formula, the outputs of all heads are concatenated, and a linear transformation is performed through the output weight matrix to obtain the final self-attention output: ; in, This represents the final self-attention output. This indicates that the attention output of the first head will be... To the Attention output of size To splice, Indicates the output weight matrix and , The dimension is The real number space.

[0043] The output of the multi-head self-attention layer, i.e. the final self-attention output, is reshaped back to the original spatial dimension and added to the output of the convolution path, enabling deep neural networks to effectively capture global dependencies in time-frequency feature maps, which is crucial for modeling complex interference patterns.

[0044] For example, the decoder provides an upsampling path consisting of four cascaded upsampling blocks. The structure of each upsampling block is similar to that of the downsampling blocks, including a ResNet layer, a SiLU activation function, and an upsampling operation layer. The ResNet layer contains two ResNetBlock2D layers, with the output of the previous ResNetBlock2D layer serving as the input to the next. The upsampling operation layer is a two-dimensional upsampling layer, UpSample2D, which achieves upsampling through transposed convolution. The input to the two-dimensional upsampling layer is the output of the last two-dimensional residual network block.

[0045] The upsampling process of the upsampling block is represented as follows: ; in, Indicates the first Decoder features of the layer Indicates the first Decoder features of the layer Indicates the first The upsampling block of the layer; This indicates a channel splicing operation implemented by a skip connection. This structure combines high-resolution features of the same scale in the encoder path with upsampled features in the decoder path, effectively conveying detailed information and ensuring the fidelity of the time-frequency structure of the output signal; when hour, The intermediate path is the output feature map of the 4th layer. The processed feature maps are as follows. The upsampled blocks of the first layer are mirror images of the downsampled blocks of the first layer. Similarly, the upsampled blocks of the second layer are mirror images of the downsampled blocks of the second layer, the upsampled blocks of the third layer are mirror images of the downsampled blocks of the third layer, and the upsampled blocks of the fourth layer are mirror images of the downsampled blocks of the fourth layer. In particular, the input of the upsampled block of the fourth layer is the feature map output by the intermediate path on the downsampled block of the fourth layer, i.e., the output feature map of the fourth layer.

[0046] The two downsampling blocks in the decoder and encoder, each containing a multi-head self-attention layer (i.e., the two attention downsampling blocks), and the two upsampling blocks in the third and fourth layers (i.e., the upsampling blocks), are mirror images of each other. Both are integrated self-attention 2D convolutional upsampling blocks (AttnUpBlocks), and both include multi-head self-attention layers. While performing transposed convolutional upsampling, the attention upsampling blocks also utilize a self-attention mechanism to fuse global contextual information. This involves combining... Figure 2 Similar to the attention downsampling block, the attention upsampling block introduces a multi-head self-attention layer between the ResNet layer and the upsampling operation layer. The input of this multi-head self-attention layer is the output of the ResNet layer, which is also the output of the last ResNetBlock2D layer in the ResNet layer. The output of this multi-head self-attention layer is the input of the upsampling operation layer, i.e., UpSample2D.

[0047] For example, the intermediate path includes ResNet layers. Specifically, the ResNet layers in the intermediate path may contain two ResNetBlock2D layers, with the output of the first ResNetBlock2D layer serving as the input of the second ResNetBlock2D layer.

[0048] In general, combining Figure 2 ,when At that time, the input to the downsampling block of the first layer is a four-channel tensor constructed using the complex time spectrum of the radar, including the complex time spectrum of the main signal and the complex time spectrum of the auxiliary signal. The output of the downsampling block of the first layer is the output feature map of the first layer. The output feature map of the first layer It is passed to the downsampling block of the second layer and the upsampling block of the first layer. When At that time, the input to the downsampling block of the second layer is the output feature map of the first layer. The output of the downsampling block of the second layer is the output feature map of the second layer. The output feature map of the second layer The samples are passed to the downsampled blocks of the third layer and the upsampled blocks of the second layer. At that time, the input of the downsampling block of the third layer is the output feature map of the second layer. The output of the downsampling block in the third layer is the output feature map of the third layer. The output feature map of the third layer The samples are passed to the downsampled blocks of the fourth layer and the upsampled blocks of the third layer. At that time, the input of the downsampling block of the fourth layer is the output feature map of the third layer. The output of the downsampling block of the fourth layer is the output feature map of the fourth layer. The output feature map of the fourth layer It is passed to the intermediate path. The input to the intermediate path is the output feature map of the fourth layer. The output of the intermediate path is the feature map of the fourth layer obtained by using the ResNet layer. Feature map after processing The input to the upsampling block of the fourth layer is the feature map output from the intermediate path. The output of the upsampled block of the fourth layer is the decoder feature of the fourth layer. The features of the fourth layer decoder The samples are passed to the upsampled blocks of the third layer. The input to the upsampled blocks of the third layer is the decoder feature of the fourth layer. and the output feature map of the third layer The output of the upsampled block of the third layer is the decoder feature of the third layer. The features of the third layer decoder The samples are passed to the upsampled blocks of the second layer. The input to the upsampled blocks of the second layer is the decoder feature of the third layer. and the output feature map of the second layer The output of the upsampled block of the second layer is the decoder feature of the second layer. The features of the second layer decoder The samples are passed to the first-layer upsampled block. The input to the first-layer upsampled block is the decoder feature of the second layer. and the output feature map of the first layer The output of the upsampled block of the first layer is the decoder feature of the first layer. The first layer decoder features This refers to a deep neural network based on a symmetric encoder-decoder architecture and integrating a self-attention mechanism. The final output is the complex time-frequency spectrum of the target signal. In the encoder, the downsampling blocks in the third and fourth layers are both attention-based downsampling blocks, and in the decoder, the upsampling blocks in the third and fourth layers are both attention-based upsampling blocks. By embedding self-attention mechanisms in the encoder and decoder, the deep neural network can efficiently compute the global dependencies between any two locations in the time-frequency feature map.

[0049] Deep Neural Networks The final output has one dimension. Predicted target time-frequency tensor , The two channels represent the real and imaginary parts of the estimated complex spectrum of the target signal, respectively. (Deep Neural Network) parameters By minimizing the time-frequency tensor of the prediction target Compared with the real target time spectrum Optimization is achieved by comparing the reconstructed reality with the actual reality. For example, deep neural networks... loss function Mean Squared Error (MSE) is used, and it is expressed as: ; in, Representing deep neural networks The parameters, Represents the number of training samples. Indicates the training sample number. Indicates the first The real target time-frequency tensor corresponding to each training sample Indicates the first The four-channel tensor corresponding to each training sample Indicates will Input deep neural network The obtained time-frequency tensor of the predicted target, This represents the L2 norm.

[0050] Specifically, when using the loss function For deep neural networks parameters When performing optimization, the optimization goal is to adjust the parameters. Make the loss function Minimize. Each training sample includes its corresponding four-channel tensor and the real target time-frequency tensor. The four-channel tensor for each training sample is constructed as follows: the main signal and auxiliary signal of the training sample are aligned; the main signal and the aligned auxiliary signal are then subjected to short-time Fourier transforms to obtain the complex time-frequency spectra of the main signal and the auxiliary signal, respectively; both the complex time-frequency spectra of the main signal and the auxiliary signal are decomposed into real and imaginary components; and these components are concatenated along the channel dimension to construct the four-channel tensor. The real target time-frequency tensor for each training sample is constructed as follows: the real target echo estimation signal of the training sample is subjected to a short-time Fourier transform to obtain the complex time-frequency spectrum of the real signal; this complex time-frequency spectrum is decomposed into real and imaginary components; and these two components are concatenated along the channel dimension to obtain the two-channel real target time-frequency tensor.

[0051] By optimizing the parameters of a deep neural network using a large number of training samples, the deep neural network can implicitly learn various patterns of disturbance from a large amount of data. It has a strong adaptive and generalization ability to the non-stationary characteristics of disturbance and environmental changes, and can cope with a variety of complex disturbance scenarios.

[0052] Step 4: Signal Reconstruction: The complex time spectrum of the target signal is reconstructed back into the time domain signal through short-time inverse Fourier transform, and the time domain signal is used as the target echo estimation signal after suppressing the interference from the same source.

[0053] For example, reconstructing the time-domain signal from the complex time spectrum of the target signal using a short-time inverse Fourier transform includes: According to the following formula, the real and imaginary parts of the complex time spectrum of the target signal are combined into a complex time-frequency matrix of the complex time spectrum of the target signal. : ; in, This represents the real part of the spectrum of the target signal when it is complex. This represents the imaginary part of the spectrum of the target signal when it is complex. Represents the imaginary unit; According to the following formula, the complex time spectrum of the target signal can be reconstructed back into the time domain signal through short-time inverse Fourier transform: ; in, The reconstructed result The time-domain signal at time t, express The short-time Fourier transform and , express The window function at time t, This indicates the transmission delay between the main channel and the auxiliary channel. Indicates frequency, Represents the natural constant.

[0054] The final time-domain signal This refers to the target echo estimation signal after suppressing interference from the same source.

[0055] The radar adaptive interference cancellation method based on a deep time-frequency network provided in this disclosure, compared with the prior art, uses a four-channel tensor of the time-frequency signal as a conditional input and directly regresses the target output through an end-to-end deep neural network based on a symmetric encoder-decoder architecture and integrating a self-attention mechanism. It does not rely on traditional iterative algorithms or explicit linear filtering models, exhibiting strong nonlinear mapping and feature separation capabilities. This deep neural network can learn and fit the complex mixing relationship between interference and echo signals in the time-frequency domain, thereby achieving more accurate interference estimation and elimination, obtaining a higher interference cancellation ratio and lower signal distortion, significantly improving the radar's detection performance in strong interference environments. Furthermore, the forward inference process of the radar adaptive interference cancellation method based on a deep time-frequency network provided in this disclosure is calculated in one step using the optimized deep neural network, resulting in fast computation speed, low latency, and easy acceleration on existing GPUs or dedicated AI chips, meeting the real-time processing requirements of radar systems.

[0056] Those skilled in the art will understand that the above embodiments are specific implementations of this disclosure, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of this disclosure.

Claims

1. A radar adaptive interference cancellation method based on deep time-frequency networks, characterized in that, The radar adaptive interference cancellation method based on deep time-frequency networks includes: The mixed signal obtained from the main radar receiver is used as the main signal, and the reference signal obtained from the auxiliary radar receiver is used as the auxiliary signal. The main signal and the auxiliary signal are then aligned to obtain the aligned auxiliary signal. The main signal and the aligned auxiliary signal are subjected to short-time Fourier transform processing respectively to obtain the complex time spectrum of the main signal and the complex time spectrum of the auxiliary signal. The complex time spectrum of the main signal and the complex time spectrum of the auxiliary signal are decomposed into two components, real part and imaginary part, and then concatenated in the channel dimension to construct a four-channel tensor. The four-channel tensor input is based on a deep neural network with a symmetric encoder-decoder architecture and an integrated self-attention mechanism to obtain the complex time spectrum of the target signal output by the deep neural network. The complex time spectrum of the target signal is reconstructed back into the time domain signal through short-time inverse Fourier transform, and the time domain signal is used as the target echo estimation signal after suppressing co-source interference.

2. The radar adaptive interference cancellation method based on deep time-frequency networks according to claim 1, characterized in that, The step of aligning the main signal and the auxiliary signal to obtain the aligned auxiliary signal includes: The main signal and the auxiliary signal are aligned using a generalized cross-correlation-phase transform weighting algorithm.

3. The radar adaptive interference cancellation method based on deep time-frequency networks according to claim 2, characterized in that, The step of using a generalized cross-correlation-phase transform weighting algorithm to perform signal alignment processing on the main signal and the auxiliary signal includes: The cross-power spectrum between the main signal and the auxiliary signal is determined according to the following formula. : ； in, The Fourier transform of the main signal, The Fourier transform of the auxiliary signal The complex conjugate; The PHAT weighting function is determined according to the following formula. : ； The weighted cross-power spectrum is determined according to the following formula. : ； According to the following formula, the weighted cross-power spectrum Performing the inverse Fourier transform yields the generalized cross-correlation function. : ； in, Indicates the inverse Fourier transform. This indicates the transmission delay between the main channel and the auxiliary channel; generalized cross-correlation function The corresponding value when the maximum value is obtained The value of As the estimated time delay, according to the following formula, using... Time delay compensation is performed on the auxiliary signal to obtain the aligned auxiliary signal. : ； in, Indicates the current moment. express Auxiliary signals for timing.

4. The radar adaptive interference cancellation method based on deep time-frequency networks according to claim 1, characterized in that, The four-channel tensor is represented as follows: ； in, For the four-channel tensor and , The dimension is The real space, Indicates the height of the time-frequency image. Indicates the width of the time-frequency image. This indicates a splicing operation. The real component of the spectrum of the main signal in its complex form is represented. This represents the imaginary part of the spectrum of the main signal when it is complex. This represents the real component of the complex frequency spectrum of the auxiliary signal. This represents the imaginary component of the spectrum of the auxiliary signal when it is complex.

5. The radar adaptive interference cancellation method based on deep time-frequency networks according to claim 1, characterized in that, The deep neural network includes an encoder and a decoder that is mirror-symmetric to the encoder, with an intermediate path provided between the encoder and the decoder.

6. The radar adaptive interference cancellation method based on deep time-frequency networks according to claim 5, characterized in that, The encoder includes four cascaded downsampling blocks, each of which contains a ResNet layer, an activation function SiLU, and a downsampling operation layer. The downsampling process of the downsampling block is represented as follows: ； in, Indicates the first The output feature map of the layer, Indicates the first The output feature map of the layer, Indicates the first The downsampling block of the layer, The values ​​are 1, 2, 3, and 4 respectively. hour, It is a four-channel tensor; The latter two downsampling blocks in the encoder also each contain a multi-head self-attention layer, the calculation process of which includes: According to the following formula, for input features Perform a linear transformation to obtain the corresponding query matrix. Key matrix Sum matrix : ； in, The dimension is The real space, Indicates channel dimension, Represents the total number of spatial locations and , Indicates the height of the time-frequency image. Indicates the width of the time-frequency image. These are the query weight matrix, key weight matrix, and value weight matrix, respectively. , The dimension is The real space, Indicates the projection dimension; query matrix Key matrix Sum matrix Divided into Each head has a dimension of [number]. For the first For each size, calculate its attention output according to the following formula: ； in, Indicates the header number and , Indicates the first Attention output based on size express function, They represent the first The corresponding query matrix, key matrix, and value matrix for each head. Indicates matrix transpose; According to the following formula, the outputs of all heads are concatenated and then linearly transformed using the output projection matrix to obtain the final self-attention output: ； in, This represents the final self-attention output. This indicates that the attention output of the first head will be... To the Attention output of size To splice, Indicates the output weight matrix and , The dimension is The real number space.

7. The radar adaptive interference cancellation method based on deep time-frequency networks according to claim 6, characterized in that, The decoder includes four cascaded upsampling blocks, and the upsampling process of each upsampling block is represented as follows: ； in, Indicates the first Decoder features of the layer Indicates the first Decoder features of the layer Indicates the first The upsampling block of the layer, This indicates a channel splicing operation implemented by skip connections, when hour, The intermediate path is the output feature map of the 4th layer. The feature map output after processing; The two upsampling blocks in the decoder that are mirror images of the two downsampling blocks in the encoder, both of which contain the multi-head self-attention layer, also each contain the multi-head self-attention layer.

8. The radar adaptive interference cancellation method based on deep time-frequency networks according to claim 5, characterized in that, The intermediate path includes a ResNet layer.

9. The radar adaptive interference cancellation method based on deep time-frequency networks according to claim 1, characterized in that, The loss function of the deep neural network Represented as: ； in, Representing deep neural networks The parameters, Represents the number of training samples. Indicates the training sample number. Indicates the first The real target time-frequency tensor corresponding to each training sample Indicates the first The four-channel tensor corresponding to each training sample Indicates will Input deep neural network The obtained time-frequency tensor of the predicted target, This represents the L2 norm.

10. The radar adaptive interference cancellation method based on deep time-frequency networks according to claim 1, characterized in that, The step of reconstructing the complex time spectrum of the target signal back to the time domain signal through short-time inverse Fourier transform includes: According to the following formula, the real and imaginary parts of the complex time spectrum of the target signal are combined into a complex time-frequency matrix of the complex time spectrum of the target signal. : ； in, This represents the real part of the spectrum of the target signal when it is complex. This represents the imaginary part of the spectrum of the target signal when it is complex. Represents the imaginary unit; According to the following formula, the complex time spectrum of the target signal can be reconstructed back into the time domain signal through short-time inverse Fourier transform: ； in, The reconstructed result The time-domain signal at time t, express The short-time Fourier transform and , express The window function at time t, This indicates the transmission delay between the main channel and the auxiliary channel. Indicates frequency, Represents the natural constant.

Images