Target signal enhancement method based on generative adversarial network

Abstract

The invention discloses a target signal enhancement method based on a generative adversarial network, solves the problems of low signal-to-noise ratio and blind signals in a complex electromagnetic environment in the prior art, and realizes detection accuracy and generalization ability in a complex background. The method comprises the following steps: setting related parameters of a receiving station, and determining a reference signal and a noisy echo signal; carrying out time-frequency transformation on the two-dimensional mutual fuzzy matrix to generate a two-dimensional mutual fuzzy matrix; extracting a distance dimension slice set of the target feature region in the two-dimensional mutual fuzzy matrix, and performing spatial sampling on each distance dimension slice to generate a noisy signal set corresponding to each distance dimension slice; inputting the noisy signal set into a pre-trained signal enhancement model to obtain an enhanced signal set corresponding to each distance dimension slice; counting the continuous correct matching times of the enhanced signal in the target position neighborhood, and judging that the target exists when the times reach a set threshold value; and if wrong matching occurs, resetting continuous counting and continuing detection until a threshold value is reached or scanning is ended.

Description

Technical Field

[0001] This invention relates to the fields of radar, communication and electronic countermeasures, and in particular to a target signal enhancement method based on generative adversarial networks. Background Technology

[0002] In radar, communications, and electronic warfare fields, target signal enhancement is a crucial step in subsequent detection, classification, and localization. However, in real-world environments, signals are often affected by multipath effects, non-Gaussian noise (such as Rayleigh noise and impulse noise), and human interference, leading to severe signal quality degradation. Traditional signal enhancement techniques mainly include the following categories: (1) Filtering methods: such as Wiener filtering and Kalman filtering, which rely on the statistical characteristics of noise and signal, and have limited effect when the noise is non-stationary or unknown.

[0003] (2) Time-frequency analysis: The signal and noise are separated by short-time Fourier transform (STFT) or wavelet transform, but the resolution is limited by the selection of time-frequency window and is sensitive to instantaneous noise.

[0004] (3) Sparse representation: The compressed sensing theory is used to recover the signal from a small number of observations, but the signal needs to be sparse and the computational complexity is high.

[0005] In recent years, deep learning technology has shown great potential in the field of signal processing, especially Generative Adversarial Networks (GANs), which have achieved remarkable results in tasks such as image denoising and speech enhancement. Through adversarial training between the generator and the discriminator, GANs can learn complex data distributions, thereby generating high-quality enhanced signals.

[0006] Kong Lingze et al. proposed a collaborative saliency detection algorithm based on generative adversarial networks and contrastive learning in the paper "Collaborative Saliency Detection Based on Generative Adversarial Networks and Contrastive Learning [J]. Jiangsu University of Science and Technology, 2025, 39(03):33-39+100.DOI:10.20061 / j.issn.1673-4807.2025.03.006". Combining positive and negative sample supervision and generative adversarial strategies, the algorithm greatly improves the model's ability to distinguish between interference objects and complex backgrounds. Xiong Hongqiang et al. proposed a collaborative saliency detection algorithm based on generative adversarial networks and contrastive learning in the paper "Ground Penetrating Radar Target Detection Method and Application Research Based on Generative Adversarial Networks [D]. Jilin University, 2024.DOI:10.2716. The paper "2 / d.cnki.gjlin.2024.000751" proposes GPR-GAN to enhance the diversity of ground-penetrating radar data and improve the generalization ability of the detection model. It also develops 2C-GAN and MinR-GAN to suppress strong clutter interference and improve the quality of low signal-to-noise ratio data. Huang Cheng et al. proposed a target detection network enhancement method based on generative adversarial architecture in the paper "Research on target detection enhancement algorithm based on generative adversarial architecture [D]. Nanjing University of Posts and Telecommunications, 2023.DOI:10.27251 / d.cnki.gnjdc.2023.000049". This method uses the data distribution fitting ability of generative adversarial architecture to improve the feature extraction ability of the network.

[0007] Traditional detection techniques cannot quickly detect low signal-to-noise ratio or blind signals in complex electromagnetic environments. Summary of the Invention

[0008] This invention provides a target signal enhancement method based on generative adversarial networks, which solves the problem in existing technologies that cannot quickly detect low signal-to-noise ratio and blind signals in complex electromagnetic environments. By introducing an adversarial learning mechanism, the network can learn the feature distribution of real radar signals while effectively suppressing noise interference, thereby enhancing the ability to extract and distinguish weak and low-visibility target signals and improving the detection accuracy and generalization ability in complex backgrounds.

[0009] This invention provides a target signal enhancement method based on generative adversarial networks, the method comprising: Set the relevant parameters of the receiving station and determine the reference signal and the noisy echo signal; A time-frequency transformation is performed on the reference signal and the noisy echo signal to generate a two-dimensional mutual ambiguity matrix. ; Extract the distance dimension slice set of the target feature region from the two-dimensional mutual blur matrix, and then extract the distance dimension slice set from each distance dimension slice. Spatial sampling is performed to generate noisy signal sets corresponding to each distance dimension slice; The noisy signal set is input into a pre-trained signal enhancement model to obtain the enhanced signal set corresponding to each distance dimension slice; The presence of a target is determined by counting the number of consecutive correct matches of the enhanced signal in the neighborhood of the target location. If the number of matches reaches a set threshold, the consecutive count is reset and detection continues until the threshold is reached or the scan is terminated.

[0010] In one possible implementation, the noisy echo signal is represented as: ; in, Indicates the attenuation coefficient; Indicates a reference signal; Indicates time delay; Indicates the Doppler frequency; Indicates noise; This indicates a noisy echo signal.

[0011] In one possible implementation, the two-dimensional mutually fuzzy matrix is ​​represented as: ; in, Indicates a reference signal; Indicates a noisy echo signal; Indicates time delay; Indicates the Doppler frequency; This represents a two-dimensional mutually fuzzy matrix.

[0012] In one possible implementation, the slices in each distance dimension Spatial sampling is performed to generate noisy signal sets corresponding to each distance dimension slice, including: At the distance from the peak position of the target dimension Place, with long windows Step Perform a rectangular sliding window to generate a set of local signal segments; Based on the local enhancement mode of the target peak, the set of local signal segments is divided into 128 label categories; For each category, generate a set of noisy signals corresponding to each distance dimension slice.

[0013] In one possible implementation, the set of local signal segments is represented as: ; in, Indicates the number of noisy signals; Indicates the stepping of a rectangular sliding window; Indicates the length of a rectangular sliding window; Represents a slice in the distance dimension; This represents a set of noisy signals.

[0014] In one possible implementation, the training process of the signal enhancement model includes: Determine the sample reference signal and the sample noisy echo signal; A time-frequency transformation is performed on the sample reference signal and the sample noisy echo signal to generate a sample two-dimensional mutual ambiguity matrix. ; Extract sample distance dimension slices of target feature regions from the sample two-dimensional mutual blur matrix, and perform spatial sampling on the sample distance dimension slices to generate a sample noisy signal set; At the target peak position of the distance dimension of the sample distance dimension slice, a rectangular sliding window is generated with a fixed window length and step to generate a set of local signal segments of the sample; Based on the local enhancement mode of the target peak, the set of local signal segments of the sample is divided into 128 label categories; For each category, generate 256 clean samples and generate 256 noisy samples for each category under various signal-to-noise ratios; Gaussian noise with different signal-to-noise ratios was added to the distance dimension slices of each sample, generating 256 noisy samples and corresponding clean labels for each label category; The noisy samples are gradually added with noise through a forward diffusion process to obtain a signal set with gradually increasing noise. ; The signal set with gradually increasing noise is gradually denoised using a denoising network to obtain a preliminary enhanced signal; The initial enhanced signal is input into a generative adversarial network (GAN), and the network parameters are optimized through adversarial training of the generator G and the discriminator D to obtain a pre-trained signal enhancement model.

[0015] In one possible implementation, the generator adopts a U-Net structure and embeds an SE-Net channel attention mechanism, while the discriminator adopts a multi-layer convolutional structure combined with a Patch discriminative structure. The generator parameter update rule is as follows: ; The discriminator parameter update rule is as follows: ; in, Indicates the learnable parameters of the generator; The gradient operator of the generator; Represents the generator loss function; These represent the learnable parameters of the discriminator; Indicates the learning rate; This represents the discriminator loss function; This represents the gradient operator of the discriminator.

[0016] In one possible implementation, determining the cumulative number of consecutive correct matches in the spatial location distribution of the enhanced signal set, determining that the target exists when the number of matches reaches a set threshold, and resetting the count in case of an incorrect match, includes: Initialize the consecutive correct match counter and the consecutive error counter, and set the fault tolerance range and fault tolerance limit; Traverse each distance dimension slice and detect whether there are sampling points within the fault tolerance range where the intensity of the enhanced signal is greater than a preset threshold; If it exists, the consecutive correct match counter is counted, and the consecutive error counter is reset to 0; If it does not exist, the continuous error counter counts. If the current continuous error counter count is greater than or equal to the fault tolerance limit, the continuous correct match counter and the continuous error counter are reset. If the continuous error counter count is less than the fault tolerance limit, the continuous correct match counter and the continuous error counter remain unchanged. After each update of the continuous correct match counter, it is determined in real time whether the value of the continuous correct match counter has reached the pre-designed value. If it has, it is determined that the target exists and the detection process is terminated.

[0017] In one possible implementation, the target feature region is determined by a preset Doppler frequency range, which is calculated based on radar motion parameters.

[0018] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention generates a two-dimensional mutually fuzzy matrix through time-frequency transformation, which effectively preserves the time and frequency characteristics of the signal, providing richer information dimensions for subsequent processing. The spatial sampling method using distance-dimensional slice sets can selectively extract target-related signal regions, reducing the amount of invalid data processing. The pre-trained signal enhancement model significantly improves the signal-to-noise ratio of noisy signals, enhancing the detection capability of weak targets. Fourth, the judgment mechanism based on the number of consecutive correct matches ensures detection reliability by setting threshold conditions, while avoiding the accumulation of misjudgments through a count reset mechanism during incorrect matches, giving the system adaptive error correction capabilities. The entire processing flow forms a closed-loop detection, ensuring real-time performance while possessing high detection accuracy and anti-interference capabilities. Attached Figure Description

[0019] Figure 1 A flowchart illustrating the steps of a target signal enhancement method based on generative adversarial networks provided in an embodiment of the present invention; Figure 2 The mutually fuzzy matrix diagram provided in the embodiments of the present invention; Figure 3 A flowchart of the diffusion model provided in an embodiment of the present invention; Figure 4 A generator structure diagram provided for an embodiment of the present invention; Figure 5 This is a diagram of the SE-Net structure provided in an embodiment of the present invention.

[0020] Figure 6 This is a structural diagram of the discriminator provided in an embodiment of the present invention.

[0021] Figure 7 These are slice images of different distance dimensions under different noise levels provided in embodiments of the present invention; Figure 8 The images provided in this embodiment of the invention are comparison images of the enhanced signal before and after at different target locations; Figure 9 A comparison chart of the number of successful enhancements of two Doppler slices provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] This invention provides a target signal enhancement method based on generative adversarial networks, see [link to relevant documentation]. Figure 1 The method includes the following steps S101 to S105.

[0024] S101, Set the relevant parameters of the receiving station and determine the reference signal and noisy echo signal; Specifically, in S101, the noisy echo signal is represented as: (1.1) in, Indicates the attenuation coefficient; Indicates a reference signal; Indicates time delay; Indicates the Doppler frequency; Indicates noise; This indicates a noisy echo signal.

[0025] For example, setting the receiving station parameters, including the sampling frequency. Signal bandwidth Target distance range The received reference signal and the noisy echo signal are represented by formula (1.1).

[0026] S102, perform time-frequency transformation on the reference signal and the noisy echo signal to generate a two-dimensional mutual ambiguity matrix. Two-dimensional mutually fuzzy matrix Please see Figure 2 .

[0027] Specifically, in S102, the two-dimensional mutual fuzzy matrix is ​​represented as: (1.2) in, Indicates a reference signal; Indicates a noisy echo signal; Indicates time delay; Indicates the Doppler frequency; This represents a two-dimensional mutually fuzzy matrix.

[0028] S103, extract the distance dimension slice set of the target feature region from the two-dimensional mutual fuzzy matrix, and then extract the distance dimension slice set from each distance dimension slice. Spatial sampling is performed to generate noisy signal sets corresponding to each distance dimension slice; Specifically, in step S103, slices are made in each distance dimension. Spatial sampling is performed to generate a set of noisy signals corresponding to each distance dimension slice, including the following steps S1031 to S1033.

[0029] S1031, at the distance from the target peak position Place, with long windows Step Perform a rectangular sliding window to generate a set of local signal segments; Here, the set of local signal segments is represented as: (1.3) in, Indicates the number of noisy signals; Indicates the stepping of a rectangular sliding window; Indicates the length of a rectangular sliding window; Represents a slice in the distance dimension; This represents a set of noisy signals.

[0030] S1032, based on the local enhancement mode of the target peak, the set of local signal segments is divided into 128 label categories; S1033, for each category, generate a set of noisy signals corresponding to each distance dimension slice, that is, generate 256 sets of noisy signals under various signal-to-noise ratios.

[0031] For example, at the target peak position in the distance dimension Place, with long windows Step A rectangular sliding window is used to generate a set of local signal segments, which is represented by formula (1.3); A total of 128 label categories were generated, each of which is a signal segment generated by the sliding window process. Each category corresponds to a different local enhancement mode of the target peak, and the set of local signal segments is divided into 128 label categories. For each Doppler slice, Gaussian noise with different signal-to-noise ratios is added to it, and 256 noisy signal sets are generated under each label category corresponding to each signal-to-noise ratio.

[0032] See Figure 7 Slices of each distance dimension under different noise levels . Figure 7 The vertical axis represents the amplitude, and the horizontal axis represents the time point.

[0033] S104, input the noisy signal set into the pre-trained signal enhancement model to obtain the enhanced signal set corresponding to each distance dimension slice; see also Figure 8 This is a comparison image of the enhanced signal before and after at different target locations. Figure 8 The vertical axis represents the amplitude, and the horizontal axis represents the sampling points.

[0034] Specifically, in step S104, the training process of the signal enhancement model includes the following steps: (1) Determine the sample reference signal and the sample noisy echo signal; (2) Perform time-frequency transformation on the sample reference signal and the sample noisy echo signal to generate a two-dimensional mutual ambiguity matrix of the samples. ; (3) Extract the sample distance dimension slice of the target feature region from the sample two-dimensional mutual fuzzy matrix, and perform spatial sampling on the sample distance dimension slice to generate a sample noisy signal set; Here, the target feature region is determined by a preset Doppler frequency range, which is calculated based on radar motion parameters.

[0035] (4) At the target peak position of the distance dimension of the sample distance dimension slice, a rectangular sliding window is generated with a fixed window length and step to generate a set of local signal segments of the sample; (5) Based on the local enhancement mode of the target peak, the set of local signal segments of the sample is divided into 128 label categories; (6) For each category, generate 256 clean samples and generate 256 noisy samples for each category under various signal-to-noise ratios; (7) Add Gaussian noise with different signal-to-noise ratios to the distance dimension slices of each sample, and generate 256 noisy samples and corresponding clean labels for each label category; (8) Noise is gradually added to the noisy sample through a forward diffusion process to obtain a signal set with gradually increasing noise. ; (9) The signal set with gradually increasing noise is gradually denoised by a denoising network to obtain a preliminary enhanced signal; (10) Input the initial enhanced signal into the generative adversarial network, and optimize the network parameters through adversarial training of the generator G and the discriminator D to obtain the pre-trained signal enhancement model.

[0036] Here, the generator adopts the U-Net structure and embeds the SE-Net channel attention mechanism, while the discriminator adopts a multi-layer convolution combined with a Patch discriminative structure; The generator parameter update rule is as follows: (1.4) The discriminator parameter update rule is as follows: (1.5) in, Indicates the learnable parameters of the generator; The gradient operator of the generator; Represents the generator loss function; These represent the learnable parameters of the discriminator; Indicates the learning rate; This represents the discriminator loss function; This represents the gradient operator of the discriminator.

[0037] For example, a total of 128 label categories are generated, where each label category is a signal segment generated by the sliding window process, and each category corresponds to a different local enhancement mode of the target peak, thus constructing a training dataset. .

[0038] Here It refers to the noisy data generated after adding noise at different signal-to-noise ratios. I set the target signal-to-noise ratio to 100dB or 1000dB (as long as the signal-to-noise ratio is large enough), and the amplitude of the target is much larger than the amplitude of the noise, so that the peak value of the target can be clearly seen.

[0039] Noisy data: Gaussian noise with different signal-to-noise ratios was added to each Doppler slice, and 256 samples were generated for each label category corresponding to each signal-to-noise ratio.

[0040] Clean data: Generate corresponding clean samples as labels, and generate 256 samples for each label category corresponding to each signal-to-noise ratio.

[0041] For noisy input signals Gaussian noise is gradually added to generate a series of signals with progressively increasing noise. Where T is the number of diffusion steps. The diffusion process is defined as: (1.6) in, It is the identity matrix. These are noise scheduling parameters that control the rate at which noise is added.

[0042] Training a noise reduction network Predict the noise added at each step, and generate an initially enhanced signal through progressive denoising. Ultimately, only the preliminary enhanced signal after preprocessing through the diffusion model will be used. The data is fed into the generator as input to generate the data. See also Figure 3 This is a schematic diagram of the diffusion model.

[0043] The denoising process is achieved by minimizing the following loss function: (1.7) in, This is real noise. The time step for random sampling.

[0044] See Figure 4 The generator employs a U-Net structure and embeds an SE-Net channel attention mechanism. The generator's input is the initially enhanced signal (i.e., the denoised signal) after preprocessing using a diffusion model. The output is the generated high-quality target signal (i.e., the enhanced signal). The generator further optimizes the input signal using the U-Net structure and the attention mechanism (SE-Net), see [link to documentation]. Figure 5 The attention mechanism (SE-Net) is used to extract key features and suppress noise, ultimately outputting a clean target signal.

[0045] See Figure 6 The discriminator distinguishes between generated signals and real signals through multi-layer convolution and patch discrimination mechanism, guiding the training of the generator. The input is the enhanced signal output by the generator or the real clean signal, and the output is the discrimination result (true or false), which is used to evaluate the quality of the generated signal.

[0046] The generator and discriminator are dynamically optimized through adversarial training. The generator attempts to generate more realistic signals to deceive the discriminator, while the discriminator continuously improves its discrimination ability. During training, the generator gradually learns the ability to recover high-quality target signals from noisy signals.

[0047] The update rules for the generator and discriminator are Equations (1.4) and (1.5), respectively.

[0048] S105: By counting the number of consecutive correct matches of the enhanced signal in the neighborhood of the target location, the target is determined to exist when the number reaches a set threshold; if an incorrect match occurs, the consecutive count is reset and detection continues until the threshold is reached or the scan is terminated.

[0049] Specifically, in step S105, the cumulative number of consecutive correct matches in the spatial location distribution of the enhanced signal set is determined, and the existence of the target is determined when the number of matches reaches a set threshold. The number of matches is reset when there is an incorrect match, including the following steps S1051 to S1054.

[0050] S1051, initialize the consecutive correct match counter and the consecutive error counter, and set the fault tolerance range and fault tolerance limit; S1052, Traverse each distance dimension slice and detect whether there are sampling points within the fault tolerance range where the intensity of the enhanced signal is greater than the preset threshold; S1053, if it exists, the continuous correct match counter counts and the continuous error counter is reset to 0; if it does not exist, the continuous error counter counts. If the current continuous error counter count is greater than or equal to the fault tolerance limit, the continuous correct match counter and the continuous error counter are reset. If the continuous error counter count is less than the fault tolerance limit, the continuous correct match counter and the continuous error counter remain unchanged. S1054: After each update of the continuous correct match counter, it is determined in real time whether the value of the continuous correct match counter has reached the pre-designed value. If it has, the target is determined to exist and the detection process is terminated.

[0051] For example, initialize parameter settings and set a continuously correct enhancement threshold. Set the maximum number of fault tolerance attempts (fault tolerance range). Initialize the consecutive correct match counter: (1.8) 128 location slices (indexes) generated for each Doppler slice ), in position At a certain point, there is a tolerance for one sampling point on the left and right. If the enhanced signal is greater than 13dB, it is considered that the consecutive correct match counter has been used once.

[0052] Continuous correct match counter update counter: (1.9) Error boosting (either not performed at the correct location or performed at the wrong location) updates the consecutive error counter: (1.10) Fault tolerance and success determination: like :maintain constant; like Reset ; When satisfied At that time, the target detection was successful.

[0053] In a specific embodiment of the present invention, in order to verify the target detection effect of the method of the present invention, Table 1 lists the detailed parameters of the experiment and performs the verification.

[0054] Table 1 Experimental parameter settings for target detection based on GAN

[0055] Figure 9 The image shows a comparison of the test results for two Doppler slices. In the -2269.449Hz Doppler slice, all 128 sliding window positions achieved continuous and correct enhancement, indicating successful target identification. However, the slice with a deviation of 10Hz (-2259.449Hz) did not reach the threshold, and the target was determined not to exist.

[0056] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0057] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. 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 therein. 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 present invention.

Claims

1. A target signal enhancement method based on generative adversarial networks, characterized in that, include: Set the relevant parameters of the receiving station and determine the reference signal and the noisy echo signal; A time-frequency transformation is performed on the reference signal and the noisy echo signal to generate a two-dimensional mutual ambiguity matrix. ; Extract the distance dimension slice set of the target feature region from the two-dimensional mutual blur matrix, and then extract the distance dimension slice set from each distance dimension slice. Spatial sampling is performed to generate noisy signal sets corresponding to each distance dimension slice; The noisy signal set is input into a pre-trained signal enhancement model to obtain the enhanced signal set corresponding to each distance dimension slice; The presence of a target is determined by counting the number of consecutive correct matches of the enhanced signal in the neighborhood of the target location. If the number of matches reaches a set threshold, the consecutive count is reset and detection continues until the threshold is reached or the scan is terminated.

2. The target signal enhancement method based on generative adversarial networks according to claim 1, characterized in that, The noisy echo signal is represented as: ； in, Indicates the attenuation coefficient; Indicates a reference signal; Indicates time delay; Indicates the Doppler frequency; Indicates noise; This indicates a noisy echo signal.

3. The target signal enhancement method based on generative adversarial networks according to claim 1, characterized in that, The two-dimensional mutually fuzzy matrix is ​​represented as: ； in, Indicates a reference signal; This indicates a noisy echo signal; Indicates time delay; Indicates the Doppler frequency; This represents a two-dimensional mutually fuzzy matrix.

4. The target signal enhancement method based on generative adversarial networks according to claim 1, characterized in that, The slice in each distance dimension Spatial sampling is performed to generate noisy signal sets corresponding to each distance dimension slice, including: At the distance from the peak position of the target dimension Place, with long windows Step Perform a rectangular sliding window to generate a set of local signal segments; Based on the local enhancement mode of the target peak, the set of local signal segments is divided into 128 label categories; For each category, generate a set of noisy signals corresponding to each distance dimension slice.

5. The target signal enhancement method based on generative adversarial networks according to claim 4, characterized in that, The set of local signal segments is represented as follows: ； in, Indicates the number of noisy signals; Indicates the stepping of a rectangular sliding window; Indicates the length of a rectangular sliding window; Represents a slice in the distance dimension; This represents a set of noisy signals.

6. The target signal enhancement method based on generative adversarial networks according to claim 1, characterized in that, The training process of the signal enhancement model includes: Determine the sample reference signal and the sample noisy echo signal; A time-frequency transformation is performed on the sample reference signal and the sample noisy echo signal to generate a sample two-dimensional mutual ambiguity matrix. ; Extract sample distance dimension slices of target feature regions from the sample two-dimensional mutual blur matrix, and perform spatial sampling on the sample distance dimension slices to generate a sample noisy signal set; At the target peak position of the distance dimension of the sample distance dimension slice, a rectangular sliding window is generated with a fixed window length and step to generate a set of local signal segments of the sample; Based on the local enhancement mode of the target peak, the set of local signal segments of the sample is divided into 128 label categories; For each category, generate 256 clean samples and generate 256 noisy samples for each category under various signal-to-noise ratios; Gaussian noise with different signal-to-noise ratios was added to the distance dimension slices of each sample, generating 256 noisy samples and corresponding clean labels for each label category; The noisy samples are gradually added with noise through a forward diffusion process to obtain a signal set with gradually increasing noise. ; The signal set with gradually increasing noise is gradually denoised using a denoising network to obtain a preliminary enhanced signal; The initial enhanced signal is input into a generative adversarial network (GAN), and the network parameters are optimized through adversarial training of the generator G and the discriminator D to obtain a pre-trained signal enhancement model.

7. The target signal enhancement method based on generative adversarial networks according to claim 6, characterized in that, The generator adopts a U-Net structure and embeds an SE-Net channel attention mechanism, while the discriminator adopts a multi-layer convolution combined with a Patch discriminative structure. The generator parameter update rule is as follows: ； The discriminator parameter update rule is as follows: ； in, Indicates the learnable parameters of the generator; The gradient operator of the generator; Represents the generator loss function; These represent the learnable parameters of the discriminator; Indicates the learning rate; This represents the discriminator loss function; This represents the gradient operator of the discriminator.

8. The target signal enhancement method based on generative adversarial networks according to claim 1, characterized in that, The step of determining the cumulative number of consecutive correct matches in the spatial location distribution of the enhanced signal set, determining that the target exists when the number of matches reaches a set threshold, and resetting the count in case of an incorrect match includes: Initialize the consecutive correct match counter and the consecutive error counter, and set the fault tolerance range and fault tolerance limit; Traverse each distance dimension slice and detect whether there are sampling points within the fault tolerance range where the intensity of the enhanced signal is greater than a preset threshold; If it exists, the consecutive correct match counter is counted, and the consecutive error counter is reset to 0; If it does not exist, the continuous error counter counts. If the current continuous error counter count is greater than or equal to the fault tolerance limit, the continuous correct match counter and the continuous error counter are reset. If the continuous error counter count is less than the fault tolerance limit, the continuous correct match counter and the continuous error counter remain unchanged. After each update of the continuous correct match counter, it is determined in real time whether the value of the continuous correct match counter has reached the pre-designed value. If it has, it is determined that the target exists and the detection process is terminated.

9. The target signal enhancement method based on generative adversarial networks according to claim 1, characterized in that, The target feature region is determined by a preset Doppler frequency range, which is calculated based on radar motion parameters.

Images