Radar composite interference semantic detection and multi-frame joint signal reconstruction method

Abstract

This invention discloses a method for semantic detection and multi-frame joint signal reconstruction of radar composite interference, belonging to the field of radar signal processing technology. For signal reconstruction, it includes performing short-time Fourier transform and power-law nonlinear transform on mixed and clean interference signals to obtain enhanced time-frequency maps and binary masks; constructing a U-shaped network guided by physical features, embedding two-dimensional constant false alarm rate (CFAR) detection in skip connections, generating a matrix of interference signal-to-noise ratio (SNR), confidence level, and neighborhood interference proportion as physical features, and outputting a semantic mask; using the CLEAN algorithm to suppress interference, and achieving signal reconstruction based on a corrected dictionary; and solving for sparse vectors in the Doppler domain through sparse Bayesian learning to obtain a high-resolution two-dimensional distribution of the target in the range and Doppler domains. This invention improves the accuracy and anti-interference robustness of high-resolution two-dimensional reconstruction of radar targets in the range and Doppler domains through U-shaped network semantic segmentation guided by physical features and sparse Bayesian learning.

Description

Technical Field

[0001] This invention discloses a method for semantic detection of radar composite interference and multi-frame joint signal reconstruction, belonging to the field of radar signal processing technology. Background Technology

[0002] To further enhance radar detection capabilities against weak clustered targets under complex interference, deep learning-based interference suppression schemes have been widely applied in recent years. The core of these methods lies in mining the time-frequency structure and morphological features of signals to distinguish between rapidly changing composite interference and target echoes, thereby achieving interference suppression and target recovery. One type of method is based on semantic segmentation or feature classification, using attention-enhanced networks and fusing time-frequency analysis with traditional classifiers to locate and zero out interference regions in the time-frequency domain, achieving target signal recovery. Another type is based on feature separation methods such as generative adversarial training, unsupervised learning, or complex numerical encoding / decoding, adaptively mining deep differences between targets and interference to adapt to unknown composite interference scenarios and ensure the integrity of target signal amplitude and phase information. A third type is based on dual-model fusion, locating interference regions and extracting multi-dimensional features of the remaining signals to achieve interference suppression and extraction of true target information. However, existing deep learning methods still suffer from insufficient depth in mining the time-frequency structure and morphological features of signals, resulting in irreversible damage to the target signal's frequency band, leading to a decrease in radar resolution and an increase in interference sidelobes. At the same time, in scenarios with rapidly changing main lobe composite interference, it is difficult to accurately preserve the amplitude and phase information of the target signal simultaneously, resulting in insufficient interference suppression accuracy. Furthermore, they do not fully consider both interference suppression and target signal integrity, further increasing the difficulty of detecting and locating weak cluster targets under low SJR conditions. Summary of the Invention

[0003] The purpose of this invention is to provide a method for semantic detection of radar composite interference and multi-frame joint signal reconstruction, in order to solve the problems in the prior art, such as the difficulty in effectively separating interference and target echo in radar under composite interference environment, the low accuracy of target reconstruction after traditional methods suppress interference, the lack of physical feature guidance in existing deep learning methods leading to inaccurate semantic segmentation masks, and the impact of dictionary mismatch on high-resolution imaging during sparse reconstruction.

[0004] A method for semantic detection of radar composite interference and joint signal reconstruction across multiple frames, including: S1. Perform short-time Fourier transform and power-law nonlinear transform on the mixed signal and the pure interference signal respectively to obtain the enhanced mixed time-frequency map and the enhanced pure interference time-frequency map. Binarize the enhanced pure interference time-frequency map to obtain the mask image. S2. Embed a physical feature extraction layer into the U-shaped network model to construct a physical feature-guided U-shaped network semantic segmentation model. Use the enhanced mixed time-frequency map as input and the mask image as the label for neural network training. After training, the final semantic mask is output. The physical feature extraction layer includes two-dimensional constant false alarm rate detection. Based on the detection results, construct the interference signal-to-noise ratio matrix, confidence matrix, and neighborhood interference ratio matrix, and perform channel concatenation to obtain the physical feature components. S3. Based on the final semantic mask, use the CLEAN algorithm to suppress interference in the mixed time-frequency matrix. The process involves processing and outputting the time-frequency matrix after interference suppression and the time-frequency domain without amplitude interference. The baseband echo signal is sampled at a fixed slow time to obtain the target discrete signal. This discrete signal is represented by a dictionary matrix and sparse vectors. Short-time Fourier transform, power-law nonlinear transform, and binarization are performed on the dictionary atoms sequentially to obtain the corresponding mask. Based on the mask and the time-frequency domain without amplitude interference, the dictionary matrix is ​​corrected to obtain the final corrected dictionary. S4: The time-frequency matrix after interference suppression undergoes inverse short-time Fourier transform, azimuth-based concatenation, and Fourier transform to obtain the transformed Doppler domain data matrix. The sparse vectors of each Doppler channel in the Doppler domain data matrix are concatenated along the channel dimension to obtain a sparse vector set. A sparse Bayesian learning algorithm is used to solve for the sparse vector set, yielding a high-resolution two-dimensional distribution of the target in the range and Doppler domains.

[0005] S1 includes, S1.1, for mixed signals The mixed time-frequency matrix is ​​obtained by performing a short-time Fourier transform. Through the An enhanced hybrid time-frequency map is obtained by performing a power-law nonlinear transformation; S1 includes S1.2, using the interference signal received by the receiver as the pure interference signal, performing a short-time Fourier transform to obtain the time-frequency matrix of the pure interference, and performing a power-law nonlinear transformation on the time-frequency matrix of the pure interference to obtain the enhanced pure interference time-frequency map; setting a fixed threshold, binarizing the pure interference time-frequency map to obtain a mask image with interference pixels of 1 and background pixels of 0.

[0006] S2 includes S2.1, a physical feature-guided U-shaped network semantic segmentation model that takes the enhanced hybrid time-frequency map as input, uses the mask image as a label, and outputs the final semantic mask. We use a weighted sum of cross-entropy loss and Dice loss as the loss function.

[0007] S2 includes S2.2, a two-dimensional constant false alarm rate (CFAR) detection method, which includes applying a sliding window to the enhanced hybrid time-frequency graph, setting a protection unit and a reference unit for each unit under test, setting the false alarm rate, and calculating threshold parameters. : ; In the formula, Number of reference units Calculate the decision threshold based on the mean and false alarm rate of the reference cells. : ; In the formula, It is a reference unit. For the index of the reference unit, , For the first One reference unit; based on The detection and decision process yields the two-dimensional constant false alarm rate (CFAR) detection results: ; In the formula, To detect interference, No interference was detected. For the time-frequency diagram, the first Line number The value of the pixel in the column. This is the row index of the time-frequency plot. This is the column index for the time-frequency plot.

[0008] S2 includes S2.3, and the result is... pixels The set is The signal-to-noise ratio of each pixel for: ; In the formula, The signal-to-noise ratio matrix is... The height of the mixed time-frequency plot, The width of the mixed time-frequency plot, It is the set of real numbers; Confidence level of each pixel for: ; In the formula, This is the confidence matrix; Centered on the current pixel Within the neighborhood of the range, also Pixel percentage for: ; In the formula, Let be the side length of the neighborhood. For indicator functions, For the mixed time-frequency diagram Line number The value of the pixel in the column. The total number of pixels in the neighborhood. This is the neighborhood interference ratio matrix; right , and Channel splicing is performed to form physical feature components. : ; In the formula, For channel dimension, , To perform channel splicing along the channel dimension.

[0009] S3 includes, S3.1, utilizing The complement is used to estimate the noise power, and an adaptive threshold is set using the noise power; the interpolation factor and peak threshold are set for... The inverse Fourier transform is performed to obtain the time-domain signal, and the time-domain signal is then combined with the interpolation factor to perform frequency-domain interpolation to obtain the interpolated hybrid time-frequency matrix. ,right Perform iterations and detect instances exceeding the peak threshold during the iteration process. Maximum peak value, based on The maximum peak value is used to construct temporal interference atoms, and these atoms are then removed from the temporal signal until... The maximum peak value is lower than the adaptive threshold or The maximum peak value is located at In addition; the final output is the time-frequency matrix after interference suppression. : ; In the formula, let for column index, for The Column; Time-frequency domain without amplitude interference for: ; ; In the formula, for Point Fast Fourier Transform, To accumulate the time-domain waveform without amplitude interference, for The List.

[0010] S3 includes, S3.2, at a fixed slow time. Below Sampling is performed to obtain the target discrete signal : ; ; In the formula, The total number of targets Index for the target , For the corresponding delayed complex amplitude, For rectangle functions, The discrete-time sampling point number. , The sampling period is For the first The echo delay of each target, The width of the transmitted pulse, It is a natural exponential function. The imaginary unit, The frequency modulation slope of the linear frequency modulated signal. For the first The reflectance coefficient of the target For carrier wavelength, For the first Radial distance of each target; Define the target discrete signal for: ; ; ; ; In the formula, It is a dictionary matrix. It is a sparse vector. For the first A dictionary atom, For the index of dictionary atoms, , The number of columns in the dictionary. The length of the echo signal, The length of the transmitted signal, for The first in Each element.

[0011] S3 includes, S3.3, and... Perform step S1.2 to obtain the mask. ; when hour, ;when hour: ; ; In the formula, This is the inverse short-time Fourier transform. It is a short-time Fourier transform. for and Interference components with overlapping time and frequency regions For element-wise multiplication, For the first One revised dictionary atom; right Each atom in the dictionary is modified to obtain the final modified dictionary. : .

[0012] S4 includes, S4.1, and The signal is transformed into a multi-frame time-domain signal through inverse short-time Fourier transform. ,Will By splicing along the azimuth dimension, the spliced ​​time-domain signal is obtained. ,Will Perform a Fourier transform along the azimuth dimension to convert to the Doppler dimension, and obtain the transformed Doppler domain data matrix. : ; ; In the formula, The number of pulses. For the first Doppler channel observation vectors, For the pulse index, , This is a superposition of the dictionaries for each Doppler channel. For the first The sparse vectors corresponding to each Doppler channel For the first Additive noise vectors for each Doppler channel.

[0013] S4 includes S4.2, which involves concatenating the sparse vectors of each Doppler channel along the channel dimension to construct a sparse vector set. : ; Solving using the sparse Bayesian learning algorithm This yields a high-resolution two-dimensional distribution of the target in the range and Doppler domains.

[0014] Compared with existing technologies, this invention has the following advantages: This invention uses a U-shaped network semantic segmentation model guided by physical features, embedding the interference signal-to-noise ratio, confidence level, and neighborhood interference ratio matrix obtained from two-dimensional constant false alarm rate detection into the skip connections. This enables precise generation of semantic masks, achieving high-precision separation of interference and target echoes, and significantly improving interference suppression capabilities in complex electromagnetic environments. Combining the CLEAN algorithm with a signal reconstruction strategy based on a modified dictionary effectively overcomes the target energy loss and dictionary mismatch problems after interference suppression in traditional methods, ensuring the complete preservation of the target's discrete signal. Employing sparse Bayesian learning to solve for sparse vectors in the Doppler domain allows for stable acquisition of high-resolution two-dimensional distributions of the target in the range-Doppler domain, resulting in high imaging clarity and strong anti-interference robustness, making it suitable for radar target detection and identification under strong interference conditions. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 These are the training and validation loss curves of the traditional U-Net and the method of this invention; Figure 3 These are the IoU results from the three detection methods; Figure 4 These are the mIoU curves of the method of the present invention under different SNR-JSR conditions; Figure 5 It is the RD distribution of the pure echo signal; Figure 6 This is the RD distribution after interference suppression by the method of the present invention with a JSR of 15dB; Figure 7 This is the RD distribution after interference suppression by the method of the present invention with a JSR of 20dB; Figure 8 This is the RD distribution after interference suppression by the method of the present invention with a JSR of 25dB; Figure 9 This is the RD distribution after interference suppression by the method of this invention at a JSR of 30dB; Figure 10 It is the RD distribution after interference suppression by the time-frequency domain interference nulling method at a JSR of 15dB; Figure 11 It is the RD distribution after interference suppression by the time-frequency domain interference nulling method at a JSR of 20dB; Figure 12 It is the RD distribution after interference suppression by the time-frequency domain interference nulling method at a JSR of 25dB; Figure 13It is the RD distribution after interference suppression by the time-frequency domain interference nulling method at a JSR of 30dB; Figure 14 It is the RD distribution after interference suppression by the time-frequency domain interference filtering method at a JSR of 15dB; Figure 15 It is the RD distribution after interference suppression by the time-frequency domain interference filtering method at a JSR of 20dB; Figure 16 It is the RD distribution after interference suppression by the time-frequency domain interference filtering method at a JSR of 25dB; Figure 17 It is the RD distribution after interference suppression by the time-frequency domain interference filtering method at a JSR of 30dB; Figure 18 It refers to the accuracy of target detection. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0017] A method for semantic detection of radar composite interference and joint signal reconstruction across multiple frames, including: S1. Perform short-time Fourier transform and power-law nonlinear transform on the mixed signal and the pure interference signal respectively to obtain the enhanced mixed time-frequency map and the enhanced pure interference time-frequency map. Binarize the enhanced pure interference time-frequency map to obtain the mask image. S2. Embed a physical feature extraction layer into the U-shaped network model to construct a physical feature-guided U-shaped network semantic segmentation model. Use the enhanced mixed time-frequency map as input and the mask image as the label for neural network training. After training, the final semantic mask is output. The physical feature extraction layer includes two-dimensional constant false alarm rate detection. Based on the detection results, construct the interference signal-to-noise ratio matrix, confidence matrix, and neighborhood interference ratio matrix, and perform channel concatenation to obtain the physical feature components. S3. Based on the final semantic mask, use the CLEAN algorithm to suppress interference in the mixed time-frequency matrix. The process involves processing and outputting the time-frequency matrix after interference suppression and the time-frequency domain without amplitude interference. The baseband echo signal is sampled at a fixed slow time to obtain the target discrete signal. This discrete signal is represented by a dictionary matrix and sparse vectors. Short-time Fourier transform, power-law nonlinear transform, and binarization are performed on the dictionary atoms sequentially to obtain the corresponding mask. Based on the mask and the time-frequency domain without amplitude interference, the dictionary matrix is ​​corrected to obtain the final corrected dictionary. S4: The time-frequency matrix after interference suppression undergoes inverse short-time Fourier transform, azimuth-based concatenation, and Fourier transform to obtain the transformed Doppler domain data matrix. The sparse vectors of each Doppler channel in the Doppler domain data matrix are concatenated along the channel dimension to obtain a sparse vector set. A sparse Bayesian learning algorithm is used to solve for the sparse vector set, yielding a high-resolution two-dimensional distribution of the target in the range and Doppler domains.

[0018] S1 includes, S1.1, for mixed signals The mixed time-frequency matrix is ​​obtained by performing a short-time Fourier transform. Through the An enhanced hybrid time-frequency map is obtained by performing a power-law nonlinear transformation; S1 includes S1.2, using the interference signal received by the receiver as the pure interference signal, performing a short-time Fourier transform to obtain the time-frequency matrix of the pure interference, and performing a power-law nonlinear transformation on the time-frequency matrix of the pure interference to obtain the enhanced pure interference time-frequency map; setting a fixed threshold, binarizing the pure interference time-frequency map to obtain a mask image with interference pixels of 1 and background pixels of 0.

[0019] S2 includes S2.1, a physical feature-guided U-shaped network semantic segmentation model that takes the enhanced hybrid time-frequency map as input, uses the mask image as a label, and outputs the final semantic mask. We use a weighted sum of cross-entropy loss and Dice loss as the loss function.

[0020] S2 includes S2.2, a two-dimensional constant false alarm rate (CFAR) detection method, which includes applying a sliding window to the enhanced hybrid time-frequency graph, setting a protection unit and a reference unit for each unit under test, setting the false alarm rate, and calculating threshold parameters. : ; In the formula, Number of reference units Calculate the decision threshold based on the mean and false alarm rate of the reference cells. : ; In the formula, It is a reference unit. For the index of the reference unit, , For the first One reference unit; based on The detection and decision process yields the two-dimensional constant false alarm rate (CFAR) detection results: ; In the formula, To detect interference, No interference was detected. For the time-frequency diagram, the first Line number The value of the pixel in the column. This is the row index of the time-frequency plot. This is the column index for the time-frequency plot.

[0021] S2 includes S2.3, and the result is... pixels The set is The signal-to-noise ratio of each pixel for: ; In the formula, The signal-to-noise ratio matrix is... The height of the mixed time-frequency plot, The width of the mixed time-frequency plot, It is the set of real numbers; Confidence level of each pixel for: ; In the formula, This is the confidence matrix; Centered on the current pixel Within the neighborhood of the range, also Pixel percentage for: ; In the formula, Let be the side length of the neighborhood. For indicator functions, For the mixed time-frequency diagram Line number The value of the pixel in the column. The total number of pixels in the neighborhood. This is the neighborhood interference ratio matrix; right , and Channel splicing is performed to form physical feature components. : ; In the formula, For channel dimension, , To perform channel splicing along the channel dimension.

[0022] S3 includes, S3.1, utilizing The complement is used to estimate the noise power, and an adaptive threshold is set using the noise power; the interpolation factor and peak threshold are set for... The inverse Fourier transform is performed to obtain the time-domain signal, and the time-domain signal is then combined with the interpolation factor to perform frequency-domain interpolation to obtain the interpolated hybrid time-frequency matrix. ,right Perform iterations and detect instances exceeding the peak threshold during the iteration process. Maximum peak value, based on The maximum peak value is used to construct temporal interference atoms, and these atoms are then removed from the temporal signal until... The maximum peak value is lower than the adaptive threshold or The maximum peak value is located at In addition; the final output is the time-frequency matrix after interference suppression. : ; In the formula, let for column index, for The List; Time-frequency domain without amplitude interference for: ; ; In the formula, for Point Fast Fourier Transform, To accumulate the time-domain waveform without amplitude interference, for The List.

[0023] S3 includes, S3.2, at a fixed slow time. Below Sampling is performed to obtain the target discrete signal : ; ; In the formula, The total number of targets Index for the target , For the corresponding delayed complex amplitude, For rectangle functions, The discrete-time sampling point number. , The sampling period is For the first The echo delay of each target, The width of the transmitted pulse, It is a natural exponential function. The imaginary unit, The frequency modulation slope of the linear frequency modulated signal. For the first The reflectance coefficient of the target For carrier wavelength, For the first Radial distance of each target; Define the target discrete signal for: ; ; ; ; In the formula, It is a dictionary matrix. It is a sparse vector. For the first A dictionary atom, For the index of dictionary atoms, , The number of columns in the dictionary. The length of the echo signal, The length of the transmitted signal, for The first in Each element.

[0024] S3 includes, S3.3, and... Perform step S1.2 to obtain the mask. ; when hour, ;when hour: ; ; In the formula, This is the inverse short-time Fourier transform. It is a short-time Fourier transform. for and Interference components with overlapping time and frequency regions For element-wise multiplication, For the first One revised dictionary atom; right Each atom in the dictionary is modified to obtain the final modified dictionary. : .

[0025] S4 includes, S4.1, and The signal is transformed into a multi-frame time-domain signal through inverse short-time Fourier transform. ,Will By splicing along the azimuth dimension, the spliced ​​time-domain signal is obtained. ,Will Perform a Fourier transform along the azimuth dimension to convert to the Doppler dimension, and obtain the transformed Doppler domain data matrix. : ; ; In the formula, The number of pulses. For the first Doppler channel observation vectors, For the pulse index, , This is a superposition of the dictionaries for each Doppler channel. For the first The sparse vectors corresponding to each Doppler channel For the first Additive noise vectors for each Doppler channel.

[0026] S4 includes S4.2, which involves concatenating the sparse vectors of each Doppler channel along the channel dimension to construct a sparse vector set. : ; Solving using the sparse Bayesian learning algorithm This yields a high-resolution two-dimensional distribution of the target in the range and Doppler domains.

[0027] The preprocessing process of this invention includes obtaining the time-frequency matrix of the mixed signal using STFT (Short Time Fourier Transform): ; In the formula, To save time; when When saved as an image, the background noise is significant under conditions of low SNR and low JSR. To enhance the contrast of the interference, a time-frequency graph enhancement function is defined. And obtain the enhanced mixed time-frequency diagram: ; In the formula, The absolute value is used to scale the image to an 8-bit grayscale range, which effectively improves the energy contrast of the time-frequency image and helps the subsequent semantic segmentation model extract interference features as input data for subsequent model training.

[0028] For tag generation, the time-frequency matrix of clean interference is first obtained using STFT. ,according to Calculate the enhanced time-frequency diagram of the interference. A fixed threshold is set (in this invention, the fixed threshold is set to 0.01 times the maximum amplitude value of all pixels in the clean interference time-frequency image) and interference is extracted, with background as 0 and interference as 1. The resulting label can be represented as: ; in The indicative function is defined as follows: ; Then, the two-dimensional constant false alarm rate (CFAR) detection method is used to obtain the CFAR detection results, and physical feature components are constructed. The protection unit of this invention is set to cover the maximum energy spread range of the target, and the reference unit is selected from a sufficient number of independent samples outside the protection unit, and the false alarm rate is set. for Finally, the physical features are input into the U-Net network to guide semantic segmentation against interference.

[0029] For evaluating models, the mean Intersection over Union (mIoU) is commonly used to measure the pixel-level localization accuracy of semantic segmentation models for each type of target region. It is defined as follows: ; in For the number of categories, , , The definitions are shown in Table 1: Table 1. Parameter definitions in the confusion matrix ; Finally, the model uses cross-entropy loss and Dice loss as its loss functions. The definition of is: ; in The number of pixels in the predicted image. For the model to the first The posterior probability of each pixel being predicted as interference. This represents the true label. In binary classification tasks, to alleviate the imbalance between positive and negative samples, Dice loss is introduced. Its definition is: ; Finally, the two loss functions are weighted to obtain the loss function of the U-Net model of this invention: ; in and These are the weights of the two loss functions.

[0030] This invention's physically-guided U-shaped network semantic segmentation model, based on the original U-Net, embeds a 2D-CFAR physical feature extraction layer into the skip connections between the encoder and decoder, forming a dual-fusion structure of physical and semantic features. The backbone network is VGG16. The core consists of an encoder and a decoder, along with an additional physical feature extraction layer. The encoder extracts multi-scale semantic features and fuses them with physical features and upsampling results in the subsequent decoder, achieving fine-grained detection of interference in the time-frequency image. Its core function is an iterative structure of "convolution + downsampling." Therefore, this invention uses the convolutional segments of VGG16 as the encoder backbone. Based on the max-pooling position, the feature network is divided into 5 levels. Each level consists of several 3×3 convolutional layers and ReLU activations, and the resolution is halved at the end through max-pooling, ultimately obtaining 5 levels of semantic features from shallow to deep.

[0031] This invention's physical feature extraction layer expresses the "sliding window-statistics-constant false alarm rate" process in traditional 2D-CFAR as a differentiable convolution operator and constructs it as an independent physical feature extraction layer embedded network. Specifically, in the physical feature extraction layer, according to the window structure of CFAR, the position weights of the target unit and its protection unit corresponding to the center region of the window are reset to 0, and the position weights of the remaining background units are reset to 1, constructing a fixed two-dimensional convolution kernel. Applying this convolution to the input time-frequency map and performing normalization can obtain local background noise estimates equivalent to traditional 2D-CFAR at each location; based on this, a fixed threshold factor is combined to complete the CFAR decision, obtaining the CFAR detection result. Then, the corresponding three types of physical features, namely interference signal-to-noise ratio, confidence level, and neighborhood relationship, are calculated on the detection result, and they are concatenated into a three-channel physical feature map, explicitly injecting radar physical priors into the U-Net skip layer in an end-to-end manner.

[0032] In the decoding stage, high-level features are first bilinearly upsampled to gradually restore spatial resolution. Then, the semantic features obtained from VGG downsampling are concatenated with the three-channel physical features output from the CFAR physical feature extraction layer in the channel dimension, and preliminary fusion is completed through several convolutional layers. The fused result is then concatenated with the upsampled features from the previous layer, and channel refinement and compression are performed through 1×1 convolution to finally form the decoding output at this scale. Through this "VGG encoding + physical skip connections" approach, the network, while maintaining a lightweight overall structure and feasible training, explicitly embeds the 2D-CFAR background estimation and constant false alarm rate (CFAR) decision process into the feature stream, significantly enhancing the network's physical perception of interference and detection performance, especially showing better robustness under low signal-to-noise ratio and complex interference conditions.

[0033] The model of this invention is used to detect the time-frequency graph, resulting in a final semantic mask. Based on the mask, the time-frequency matrix is ​​processed using CLEAN, the core of which includes frequency domain analysis, peak detection, time-domain atom construction, and peak removal.

[0034] The final semantic mask is Its complement is ,in To and A matrix of all ones of the same size. Assume... , , The number of points in STFT is Then there is , It is a complex field.

[0035] The core of the CLEAN algorithm is to detect peaks in the spectrum and construct corresponding single-frequency signals for elimination in the time domain. This process is repeated until the peak value is below a threshold. In this invention, the interpolation factor is set to 8, and the peak value threshold is 0.1. If... This indicates that the current mask column has no interference, so it is directly retained. The signal in this column, that is .like This indicates that there is interference in the current mask column, and subsequent CLEAN processing is required. express The number of zeros in Central Africa; The index is the column number.

[0036] for The first one that contains interference First, use mask complement. from The clean region signal is extracted to estimate the noise power, and an adaptive detection threshold is set accordingly. : ; in This is the threshold adjustment coefficient. For the first Each threshold.

[0037] To improve the accuracy of CLEAN frequency estimation, the frequency domain signal is... Perform frequency domain interpolation. Let the interpolation factor be... First, perform an inverse Fourier transform on it, and then perform interpolation: ; in represent Point inverse Fourier transform, represent Point Fourier transform.

[0038] In each iteration, detection Exceeding the threshold The maximum peak value is denoted by its index. The corresponding complex amplitude is Based on this, temporal interference atoms are constructed. : ; right Accumulate data to obtain the amplitude-free values ​​of the current column. .

[0039] To ensure amplitude-scale matching and make the constructed atoms energy-matched with the original disturbance components, corresponding amplitudes need to be added to them. : ; Remove the atom from the current time domain signal After removing the middle part, the residual time-domain signal is obtained: ; And on conduct Point Fourier transform updates the frequency domain residuals, and peak detection and atom removal are repeated until the maximum spectral peak of the residuals is below a threshold. The final time-domain signal will be processed. Point Fourier transform yields The result after CLEAN processing Meanwhile, to ensure that CLEAN does not over-scrape the target, the CLEAN operation is terminated if the current peak index is outside the mask, avoiding damage to the target components. The final residual is then processed using FFT. Transform to the frequency domain to obtain and the time-frequency matrix Repeat the above CLEAN operation for each column to obtain the CLEAN-encoded time-frequency matrix: ; Meanwhile, the time-frequency domain form without amplitude interference can be expressed as: ; ; After completing the CLEAN interference suppression guided by the U-Net semantic segmentation mask, a clean time-frequency matrix is ​​obtained. .

[0040] After dictionary correction, although the interference is dense within a frame and suppresses the target signal, the location of the interference differs between frames. Therefore, coherent accumulation of the target signal can be achieved through Doppler transform. First, the time-frequency signal after CLEAN is converted to the time domain: ; For the multi-frame case, assuming the number of pulses is... The time-domain signals obtained by clearing each frame are spliced ​​along the azimuth dimension to obtain... : ; The signal of each frame can be represented by a dictionary as: ; in A dictionary corresponding to each frame of signal. It is a sparse vector, and the positions of its non-zero values ​​correspond to the time delay of the target.

[0041] In the slow time dimension, the Doppler term is first decomposed into: ; This part is sampled, and the Doppler term corresponding to each pulse is: ; in This is a slow sampling interval. Therefore, sparse vectors It can be characterized as: ; in It is a fixed sparse vector.

[0042] Will Perform an FFT along the azimuth dimension and convert it to Doppler dimensions: ; exist The In each channel, the signal components can be represented as: ; in is the twiddle factor of the FFT.

[0043] When satisfied Under certain conditions, the two exponential terms cancel each other out, and the phases are aligned. The signal of each Doppler channel can be simplified to: ; in This is a superposition of the dictionaries for each channel. For the first The sparse vectors corresponding to each Doppler channel, when the target velocity matches the current Doppler channel. Non-zero, the position of its non-zero element corresponds to the time delay of the target.

[0044] Concatenate the sparse vectors of each Doppler channel along the channel dimension to construct a sparse vector set. : ; in, Let be the joint sparse representation matrix of the distance-Doppler domain. The row index corresponds to the distance cell, representing the spatial position of the target; the column index corresponds to the Doppler channel, reflecting the radial velocity information of the target; the SBL algorithm is used to solve the sparse vector to obtain the high-resolution two-dimensional distribution of the target in the range-Doppler domain. SBL relies on the sparse Bayesian learning framework and uses an automatic correlation determination mechanism to achieve adaptive screening and pruning of dictionary atoms, which can achieve high-precision sparse coefficient estimation even with an overcomplete dictionary.

[0045] The following description, in conjunction with the accompanying drawings and embodiments, provides further details. The method flow of this invention is as follows: Figure 1 As shown, the cluster multi-frame echo signal is used as input, and frame-by-frame STFT is performed. Then, the physical features are input to guide the U-Net network model to perform time-frequency distribution. The time-frequency distribution result is divided into three branches. The first branch is directly input into the U-Net network. The second branch is processed by 2D-CFAR (two-dimensional constant false alarm rate detection method) and then input into the U-Net network. The U-Net outputs a mask. The time-frequency distribution result of the third branch and the mask are processed together using the CLEAN method to obtain the CLEAN-reduced time-frequency distribution. After interference, the CLEAN-reduced time-frequency distribution is optimized with the mask to obtain the LFM dictionary. The CLEAN-reduced time-frequency distribution is then subjected to inverse STFT to obtain the interference-suppressed cluster multi-frame echo signal. The LFM dictionary and the interference-suppressed cluster multi-frame echo signal are then processed by SBL to obtain the range-Doppler distribution.

[0046] In Embodiment 1 of this invention, a random combination of SNR and JSR within the range of 15-30 dB is used to construct the short-time Fourier transform time-frequency diagram of the signal and mask labels for training the invention. The network training parameters and radar parameters are shown in Table 2, and the interference parameters are shown in Table 3. Table 2 Training Parameter Settings ; Table 3 Interference Parameters .

[0047] Under the same dataset and training strategy, Example 1 uses the loss convergence curve and intersection-over-union ratio (IoU) as evaluation metrics to systematically compare the performance differences of the traditional U-Net, the improved U-Net of this invention, and 2D-CFAR in the time-frequency map interference detection task. Figure 2The training loss curves of the traditional U-Net and the improved U-Net of this invention are shown (the U-Net mentioned in the figure refers to the U-shaped network semantic segmentation model guided by the physical features of this invention). As can be seen from the figure, the training loss of both methods decreases to around 0.03 after approximately 40 iterations and enters a plateau, eventually converging to 0.028 and 0.020 respectively. Under the same number of iterations, the training loss of the improved U-Net model of this invention is consistently about 0.006 to 0.008 lower than that of the traditional U-Net. On the validation set, the validation loss of the improved U-Net method decreases more gradually, eventually approaching 0.020. In contrast, the validation loss of the traditional U-Net method performs worse under the same conditions. The results show that the prior information of the three channels—JSNR, confidence, and neighborhood relationship—generated through 2D-CFAR convolution provides explicit prior parameters for the network, significantly reducing the optimization difficulty and convergence error. This accelerates model convergence and improves the model's detection performance in interference detection tasks.

[0048] The IoU results of the traditional U-Net, the improved U-Net of this invention, and 2D-CFAR on the test set were statistically analyzed, such as... Figure 3 As shown in the figure. Specifically, the improved U-Net of this invention exhibits superior performance in both interference classification and background classification. In interference classification, the IoU is improved by 1.8% compared to the traditional U-Net, while the IoU is also slightly improved in background classification. Considering all categories, the mIoU of the method of this invention reaches 0.963, which is 1% higher than that of the traditional U-Net. The results fully verify the effectiveness of physical feature guidance in improving model performance. In contrast, 2D-CFAR performs the worst, with an mIoU of only 0.729, and only 0.529 in interference classification. The main reason is that noise and other interference factors in the time-frequency graph reduce the decision accuracy of CFAR, causing noise to be misclassified as interference. The improved U-Net of this invention, through physical feature guidance, can more accurately distinguish between interference and noise, thereby improving detection accuracy. In summary, physical feature guidance not only enhances the model's ability to identify interference targets, but also improves the model's robustness in complex environments, verifying its important value in radar interference detection tasks.

[0049] The mIoU under different SNR-JSR conditions was statistically analyzed, and the results are as follows: Figure 4As shown, the results demonstrate that the model of this invention exhibits excellent performance under different SNR-JSR conditions. With the increase of SNR and JSR, the model's mIoU value steadily improves, especially under low SNR and low JSR conditions, where the mIoU (interference cross-coupling ratio) remains above 0.93. This verifies the effectiveness of physical feature guidance in radar interference detection and provides pixel-level data support for subsequent interference suppression and reconstruction.

[0050] Finally, the prediction times of the traditional U-Net, the method of this invention, and 2D-CFAR were evaluated. The calculations were performed on 1000 time-frequency images, and the results are shown in Table 4. Table 4. Average prediction time per sheet for the three methods ; The results show that the method of the present invention requires the least amount of time for interference discrimination in the time-frequency diagram, combined with... Figure 4 As can be seen from the mIoU results, the method of the present invention performs optimally in both efficiency and accuracy.

[0051] Example 2 demonstrates interference suppression of echoes containing composite interference under multi-target conditions. Radar parameters are shown in Table 5, and interference parameters are shown in Table 3. Table 5 Radar Parameters ; Example 2 compares the time-frequency domain interference nulling method and the time-frequency domain interference filtering method with the method of this invention, performing a performance comparison analysis. The time-frequency domain interference nulling method directly nullifies the interference region in the time-frequency distribution, while the time-frequency domain interference filtering constructs a band-stop filter based on interference location to filter out the interference in the time-frequency domain. Finally, an inverse transformation is performed to the time domain to obtain the interference-suppressed signal. In the Doppler domain, the pulse-Doppler image contains the target's range and velocity information. The RD distribution of the clean echo is as follows... Figure 5 As shown.

[0052] Furthermore, using the method of this invention and two comparative methods, the RD distribution obtained after interference suppression of the signal in each frame is as follows: Figures 6 to 17As shown in the figure, the RD distribution after interference suppression shows that, under the same SNR, as the JSR increases, the result obtained by the method of this invention is completely consistent with the RD distribution of the clean target echo, and can accurately recover the target position and velocity information. In contrast, the time-frequency domain interference nulling method only performs nulling operations on the main lobe of the interference, and it is difficult to effectively suppress the influence of the interference sidelobes. Therefore, as the JSR increases, the interference effect of the interference sidelobes on target recovery gradually increases, leading to false alarms at some target positions. Although the time-frequency domain filtering method suppresses the interference sidelobes to a certain extent, there are still interference residues under high JSR conditions. These residual interferences will mask part of the target signal and reduce the target detection accuracy. In addition, the time-frequency domain filtering method may introduce additional spectral distortion when filtering out interference, further affecting the accurate estimation of target parameters. In contrast, the method of this invention can maintain high target detection accuracy and parameter estimation accuracy under high JSR conditions, showing significant advantages.

[0053] This invention Figure 6 , Figure 7 , Figure 8 and Figure 9 The figure shows the RD distribution of the method of the present invention as the interference-to-information ratio gradually increases from 15dB to 30dB. The results show that the method of the present invention can stably reconstruct the RD distribution of the target at 15dB, 20dB, 25dB and 30dB, with consistent restoration effect. Figure 10 and Figure 14 The RD distributions after interference suppression using the time-frequency domain interference nulling method and the time-frequency domain interference filtering method are presented respectively, at a JSR of 15dB. Both methods demonstrate stable recovery capability at 15dB, and their recovery effects are comparable to those of the time-frequency domain interference filtering method. Figures 6 to 9 Consistent, therefore Figure 6 , Figure 7 , Figure 8 , Figure 9 , Figure 10 and Figure 14 The accompanying figures are completely identical; however, the time-frequency domain interference zeroing method and the time-frequency domain interference filtering method gradually exhibit false alarms and missed detections in the RD distribution when boosted to 20dB, 25dB, and 30dB. This verifies that the method of the present invention has excellent anti-interference robustness under different interference-to-signal ratio scenarios.

[0054] Using the method of this invention and a comparative method, interference suppression is performed in multi-pulse conditions, and its accuracy is statistically analyzed. Accuracy is defined as... ,in For the number of targets that are correctly detected, This represents the total target number. The statistical results are as follows: Figure 18As shown in the results, under different JSR conditions, the method of this invention has a significant advantage over other comparative methods in terms of target detection accuracy after interference suppression of signals containing mixed interference. When the JSR is 15dB, the accuracy of the method of this invention reaches 100%, and the accuracy of the time-frequency domain interference zeroing method and time-frequency filtering can also reach over 90%. As the JSR gradually increases to 20dB and 25dB, the accuracy of the method of this invention remains stable at over 95%, while the accuracy of the comparative methods continues to decline, all dropping to around 75%. Under strong interference conditions with a JSR of 30dB, the accuracy of the method of this invention can still be maintained at 92%, while the accuracy of the comparative methods drops to around 70%.

[0055] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for semantic detection of radar composite interference and joint signal reconstruction of multiple frames, characterized in that, include: S1. Perform short-time Fourier transform and power-law nonlinear transform on the mixed signal and the pure interference signal respectively to obtain the enhanced mixed time-frequency map and the enhanced pure interference time-frequency map. Binarize the enhanced pure interference time-frequency map to obtain the mask image. S2. A physical feature extraction layer is embedded in the U-shaped network model to construct a U-shaped network semantic segmentation model guided by physical features. The enhanced hybrid time-frequency map is used as input, and the mask image is used as the label to train the neural network. After training, the final semantic mask is output. The physical feature extraction layer includes performing two-dimensional constant false alarm rate detection, constructing a signal-to-noise ratio matrix, a confidence matrix, and a neighborhood interference ratio matrix based on the detection results, and then performing channel splicing to obtain physical feature components; S3. Based on the final semantic mask, the CLEAN algorithm is used to suppress interference in the hybrid time-frequency matrix, outputting the time-frequency matrix after interference suppression and the time-frequency domain without amplitude interference. The baseband echo signal is sampled at a fixed slow time to obtain the target discrete signal. The target discrete signal is represented by a dictionary matrix and sparse vectors. The dictionary atoms are sequentially subjected to short-time Fourier transform, power nonlinear transform, and binarization to obtain the corresponding mask. Based on the mask and the time-frequency domain without amplitude interference, the dictionary matrix is ​​corrected to obtain the final corrected dictionary. S4. The time-frequency matrix after interference suppression is subjected to short-time inverse Fourier transform, azimuth splicing, and Fourier transform to obtain the transformed Doppler domain data matrix. The sparse vectors of each Doppler channel in the Doppler domain data matrix are spliced ​​along the channel dimension to obtain a sparse vector set. The sparse Bayesian learning algorithm is used to solve the sparse vector set to obtain the high-resolution two-dimensional distribution of the target in the range and Doppler domains.

2. The radar composite interference semantic detection and multi-frame joint signal reconstruction method according to claim 1, characterized in that, S1 includes, S1.1, for mixed signals The mixed time-frequency matrix is ​​obtained by performing a short-time Fourier transform. Through the A power-order nonlinear transformation is performed to obtain the enhanced hybrid time-frequency plot; S1 includes S1.2, using the interference signal received by the receiver as the clean interference signal, performing a short-time Fourier transform to obtain the time-frequency matrix of the clean interference, and performing a power nonlinear transformation on the time-frequency matrix of the clean interference to obtain the enhanced clean interference time-frequency diagram. By setting a fixed threshold, the pure interference time-frequency image is binarized to obtain a mask image with interference pixels of 1 and background pixels of 0.

3. The radar composite interference semantic detection and multi-frame joint signal reconstruction method according to claim 2, characterized in that, S2 includes S2.1, a physical feature-guided U-shaped network semantic segmentation model that takes the enhanced hybrid time-frequency map as input, uses the mask image as a label, and outputs the final semantic mask. We use a weighted sum of cross-entropy loss and Dice loss as the loss function.

4. The radar composite interference semantic detection and multi-frame joint signal reconstruction method according to claim 3, characterized in that, S2 includes S2.2, a two-dimensional constant false alarm rate (CFAR) detection method, which includes applying a sliding window to the enhanced hybrid time-frequency graph, setting a protection unit and a reference unit for each unit under test, setting the false alarm rate, and calculating threshold parameters. : ； In the formula, Number of reference units Calculate the decision threshold based on the mean and false alarm rate of the reference cells. : ； In the formula, It is a reference unit. For the index of the reference unit, , For the first One reference unit; based on The detection and decision process yields the two-dimensional constant false alarm rate (CFAR) detection results: ； In the formula, To detect interference, No interference was detected. The time-frequency diagram is the first Line 1 The value of the pixel in the column. This is the row index of the time-frequency plot. This is the column index for the time-frequency plot.

5. The radar composite interference semantic detection and multi-frame joint signal reconstruction method according to claim 4, characterized in that, S2 includes S2.3, and the result is... pixels The set is The signal-to-noise ratio of each pixel for: ； In the formula, The signal-to-noise ratio matrix is... The height of the mixed time-frequency plot, The width of the mixed time-frequency plot, It is the set of real numbers; Confidence level of each pixel for: ； In the formula, This is the confidence matrix; Centered on the current pixel Within the neighborhood of the range, also Pixel percentage for: ； In the formula, Let be the side length of the neighborhood. For indicator functions, For the mixed time-frequency diagram Line 1 The value of the pixel in the column. The total number of pixels in the neighborhood. This is the neighborhood interference ratio matrix; right , and Channel splicing is performed to form physical feature components. : ； In the formula, For channel dimension, , To perform channel splicing along the channel dimension.

6. The radar composite interference semantic detection and multi-frame joint signal reconstruction method according to claim 5, characterized in that, S3 includes, S3.1, utilizing The complement of the noise power is used to estimate the noise power, and an adaptive threshold is set using the noise power. Set the interpolation factor and peak threshold for... The inverse Fourier transform is performed to obtain the time-domain signal, and the time-domain signal is then combined with the interpolation factor to perform frequency-domain interpolation to obtain the interpolated hybrid time-frequency matrix. ,right Perform iterations and detect instances exceeding the peak threshold during the iteration process. Maximum peak value, based on The maximum peak value is used to construct temporal interference atoms, and these atoms are then removed from the temporal signal until... The maximum peak value is lower than the adaptive threshold or The maximum peak value is located at In addition; The final output is the time-frequency matrix after interference suppression. : ； In the formula, let for column index, for The List; Time-frequency domain without amplitude interference for: ； ； In the formula, for Point Fast Fourier Transform, To accumulate the time-domain waveform without amplitude interference, for The List.

7. The radar composite interference semantic detection and multi-frame joint signal reconstruction method according to claim 6, characterized in that, S3 includes, S3.2, at a fixed slow time. Below Sampling is performed to obtain the target discrete signal : ； ； In the formula, The total number of targets Index for the target , For the corresponding delayed complex amplitude, For rectangle functions, The discrete-time sampling point number. , The sampling period is For the first The echo delay of each target, The width of the transmitted pulse, It is a natural exponential function. The imaginary unit, The frequency modulation slope of the linear frequency modulated signal. For the first The reflectance coefficient of the target For carrier wavelength, For the first Radial distance of each target; Define the target discrete signal for: ； ； ； ； In the formula, It is a dictionary matrix. It is a sparse vector. For the first A dictionary atom, For the index of dictionary atoms, , The number of columns in the dictionary. The length of the echo signal, The length of the transmitted signal, for The first in Each element.

8. The radar composite interference semantic detection and multi-frame joint signal reconstruction method according to claim 7, characterized in that, S3 includes, S3.3, and... Perform step S1.2 to obtain the mask. ; when hour, ; when hour: ； ； In the formula, This is the inverse short-time Fourier transform. It is a short-time Fourier transform. for and Interference components with overlapping time and frequency regions For element-wise multiplication, For the first One revised dictionary atom; right Each atom in the dictionary is modified to obtain the final modified dictionary. : 。 9. The radar composite interference semantic detection and multi-frame joint signal reconstruction method according to claim 8, characterized in that, S4 includes, S4.1, and will The signal is transformed into a multi-frame time-domain signal through inverse short-time Fourier transform. ,Will By splicing along the azimuth dimension, the spliced ​​time-domain signal is obtained. ,Will Perform a Fourier transform along the azimuth dimension to convert to the Doppler dimension, and obtain the transformed Doppler domain data matrix. : ； ； In the formula, The number of pulses. For the first Doppler channel observation vectors, For the pulse index, , This is a superposition of the dictionaries for each Doppler channel. For the first The sparse vectors corresponding to each Doppler channel For the first Additive noise vectors for each Doppler channel.

10. The radar composite interference semantic detection and multi-frame joint signal reconstruction method according to claim 9, characterized in that, S4 includes S4.2, which involves concatenating the sparse vectors of each Doppler channel along the channel dimension to construct a sparse vector set. : ； Solving using the sparse Bayesian learning algorithm This yields a high-resolution two-dimensional distribution of the target in the range and Doppler domains.

Images