Radar anti-interference method and device based on cascade generative model

Abstract

The invention discloses a radar anti-interference method and device based on a cascade generative model, and relates to the technical field of radar anti-interference, and the method is characterized in that a target echo signal containing interference and a known prompt signal are inputted into a trained SoloRadar generative model; the models are respectively mapped to a compact potential space through an encoder. And performing denoising and feature extraction on the potential representation of the echo signal in the potential space by using a generator based on a diffusion model and taking the potential representation of the prompt signal as a guide condition to obtain the potential representation of a pure target signal. And finally, reconstructing the potential representation into a time domain signal through a decoder, optimizing the time domain signal through a corrector, and finally outputting a high-fidelity pure target echo signal. According to the method, direct operation on complex time domain signals is avoided, and the anti-interference and signal purification performance is remarkably improved through condition generation and cascade optimization in a potential space.

Description

Technical Field

[0001] This invention belongs to the field of radar anti-jamming technology, specifically relating to a radar anti-jamming method and device based on a cascaded generative model. Background Technology

[0002] With the increasing complexity of the electromagnetic environment on modern battlefields, radar, as a core piece of information sensing equipment, is facing unprecedented challenges in terms of survivability and performance. Deliberate electronic jamming methods are emerging one after another, from traditional suppression noise to intelligent deception jamming that is highly correlated with signal patterns. Their purpose is to saturate radar receivers, degrade the signal-to-noise ratio, and even inject false information, thereby seriously damaging the radar's detection, tracking, and identification capabilities.

[0003] Existing solutions to radar anti-jamming problems mainly fall into two categories. The first category is based on traditional signal processing methods, such as designing frequency-agile waveforms, constructing adaptive filters, and suppressing interference in the transform domain. These methods are effective against known interference with precise mathematical prior models, but their performance is significantly limited when facing interference with unknown, variable, or complex patterns. The second category comprises discriminative methods based on deep learning, which have emerged in recent years. These methods draw on successful experiences in computer vision, treating radar data as images and applying discriminative models such as convolutional neural networks for target detection or interference classification. However, when performing more sophisticated tasks such as signal separation and reconstruction, discriminative models are essentially a "scratch-and-recovery" mapping learning process, easily introducing unnatural computational artifacts into the output. Especially when facing interference with highly similar characteristics to the target signal, these methods struggle to recover the fine structure and crucial phase information of the signal with high quality. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a radar anti-jamming method and apparatus based on a cascaded generative model. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a radar anti-jamming method based on a cascaded generative model, comprising: Step 1: Acquire echo signals and prompt signals; Step 2: Input the echo signal and the prompt signal into the trained SoloRadar generative model to generate a clean target echo signal; The SoloRadar generative model compresses the echo signal and the cue signal into the latent space to obtain corresponding latent representations. Using the latent representation of the cue signal as a guiding condition, it denoises and extracts the latent representation of the echo signal to obtain the target latent representation. After decoding and correcting the target latent representation, the clean target echo signal is obtained.

[0005] This invention provides a radar anti-jamming device based on a cascaded generative model, applicable to the radar anti-jamming method based on a cascaded generative model described in any of the above embodiments, comprising: The signal acquisition unit is used to acquire echo signals and prompt signals; The generation unit is used to input the echo signal and the prompt signal into the trained SoloRadar generative model to generate a clean target echo signal; The SoloRadar generative model compresses the echo signal and the cue signal into the latent space to obtain corresponding latent representations. Using the latent representation of the cue signal as a guiding condition, it denoises and extracts the latent representation of the echo signal to obtain the target latent representation. After decoding and correcting the target latent representation, the clean target echo signal is obtained.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: The radar anti-jamming method based on the cascaded generative model of the present invention adopts a generative purification method, which eliminates the need to establish an accurate mathematical model of the interference. Through a diffusion model guided by the cue signal, it can directly generate a pure target echo signal from the potential space. After the signal is purified in the potential space, fine repair is performed, achieving high-fidelity signal recovery and solving the signal distortion problem caused by discriminative deep learning methods.

[0007] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0008] Figure 1 This is a flowchart of a radar anti-jamming method based on a cascaded generative model provided in an embodiment of the present invention; Figure 2 This is an overall architecture diagram of the SoloRadar generative model provided in this embodiment of the invention; Figure 3 This is a schematic diagram of the structure of the variational autoencoder provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the diffusion converter based on the U-Net architecture provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the guide converter provided in an embodiment of the present invention; Figure 6This is a schematic diagram of the uDiT module provided in an embodiment of the present invention; Figure 7 These are time-domain waveform comparison diagrams of the simulation experiments provided in this embodiment of the invention; Figure 8 This is a comparison diagram of the time-frequency domain waveforms of the simulation experiment provided in the embodiments of the present invention; Figure 9 This is a comparison chart of pulse compression results before and after interference processing provided in an embodiment of the present invention; Figure 10 This is a comparison diagram of the coherent results before and after interference processing provided in the embodiments of the present invention. Detailed Implementation

[0009] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a radar anti-jamming method and apparatus based on a cascaded generative model proposed in accordance with the present invention.

[0010] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0011] Firstly, embodiments of the present invention provide a radar anti-jamming method based on a cascaded generative model. The core idea of ​​this method is no longer to remove interference from mixed signals, but to directly generate a pure target echo signal from scratch based on prior clues. A SoloRadar generative model is designed, which can recover the pure target echo signal with high fidelity from any interference background.

[0012] Please see Figure 1 , Figure 1 This is a flowchart of a radar anti-jamming method based on a cascaded generative model provided by an embodiment of the present invention, such as... Figure 1 As shown, the radar anti-jamming method based on a cascaded generative model in this embodiment includes the following steps: Step 1: Acquire the echo signal and the prompt signal.

[0013] In this embodiment, the echo signal is the target echo signal received by the radar, which contains interference and noise signals, and the cue signal is the prior signal containing the target characteristics. In this embodiment, the standard transmission waveform of the radar is used.

[0014] Radar receiver at time Received echo signal It can be uniformly represented as a linear superposition of the clean target echo signal, one or more interference signals, and receiver thermal noise: ; in, It is a pure target echo signal. It comes from the superposition of one or more interferences. This represents additive white Gaussian noise in signal transmission.

[0015] Understandably, linear frequency modulated (LFM) signals, due to their excellent pulse compression characteristics and insensitivity to Doppler shift, are the most widely used in modern radar systems. Therefore, the LFM signal can be selected as the target signal for study, and its complex envelope form can be expressed as: ; in, For signal amplitude, The pulse width. For rectangular window functions, The center frequency of the signal. This represents the frequency modulation slope.

[0016] Optionally, intermittent sample-and-forward jamming (ISRJ) can be used as the jamming input. It is a typical coherent deception jamming mechanism. Its jamming mechanism involves high-speed periodic sampling and forwarding of the intercepted radar signal, thereby generating a dense series of false targets in the range dimension. Because these false targets and the echoes of real targets originate from the same source and have high coherence, traditional matched filters will identify them all as valid targets, resulting in severe confusion and deception. Its mathematical model can be expressed as: ; in, For the amplitude of the interference signal, The initial forwarding delay introduced by the jammer, For the sampling and forwarding period, Width of each sample (usually) ), This refers to the number of times the message was forwarded.

[0017] It should be noted that the cue signals guiding the target extraction process are highly flexible. They can be prior signals containing fine-grained target features (such as historical track data, simulated features, or standard radar transmission waveforms), or they can be cross-modal commands from other sensors. To simplify the target extraction task and verify the effectiveness of the core framework, easily obtainable and fully known standard transmission waveforms can be used. As a cue signal, it is intended to guide the generative model to identify and reconstruct the target component with the same characteristics from complex mixed echoes.

[0018] Step 2: Input the echo signal and cue signal into the trained SoloRadar generative model to generate a clean target echo signal.

[0019] In this embodiment, the SoloRadar generative model compresses the echo signal and the cue signal into the latent space to obtain the corresponding latent representations. The latent representation of the cue signal is used as a guiding condition to denoise and extract the latent representation of the echo signal to obtain the target latent representation. After decoding and correcting the target latent representation, a clean target echo signal is obtained.

[0020] Furthermore, the structure of the SoloRadar generative model in this embodiment will be described in detail. Please refer to [link / reference]. Figure 2 , Figure 2 This is an overall architecture diagram of the SoloRadar generative model provided in this embodiment of the invention, as shown below. Figure 2 As shown, the SoloRadar generative model includes a variational autoencoder operating in the time-frequency domain, a generator, and a corrector. The variational autoencoder includes an encoder and a decoder.

[0021] In this embodiment, the encoder performs a short-time Fourier transform on the input echo signal and cue signal to obtain the corresponding spectrogram, and uses the TF-GridNet module to compress the spectrogram into the latent space to obtain the corresponding latent representation; the generator is a diffusion transformer based on the U-Net architecture, used to denoise and extract the latent representation of the echo signal according to the guidance conditions to obtain the target latent representation, which is the latent representation of the clean target echo; the encoder uses the TF-GridNet module to reconstruct the target latent representation into the corresponding time-frequency diagram, and uses the inverse Fourier transform to restore the time-frequency diagram to the preliminary reconstructed signal; the corrector is a fractional single-step diffusion model, used to obtain the clean target echo signal based on the preliminary reconstructed signal and the echo signal.

[0022] Understandably, applying generative models directly to high-dimensional time-domain signals is not only computationally expensive but also prone to causing the model to focus on irrelevant noise details. In order to learn a latent space that can both compactly represent the signal and preserve its key physical properties, a variational autoencoder (VAE) operating in the time-frequency (TF) domain was designed and implemented, aiming to efficiently capture the inherent two-dimensional structural features of radar signals.

[0023] Please see Figure 3 , Figure 3 This is a schematic diagram of the variational autoencoder provided in an embodiment of the present invention, as shown below. Figure 3 As shown, specifically, a variational autoencoder consists of an encoder and a decoder. The encoder first processes the input I / Q time-domain signal... Applying the short-time Fourier transform (STFT), it is mapped to the time-frequency complex spectrum. ,in and These represent the number of frequency points and the number of time frames, respectively. Subsequently, multiple TF-GridNet modules are used. TF-GridNet is a powerful backbone network that, through alternating information exchange in the time and frequency dimensions, can simultaneously capture the local time-frequency patterns and global structural dependencies of a signal. This is crucial for accurately characterizing the fine features of radar signals. The encoder further maps the spectrogram into two latent vectors through this module: the mean vector and the mean vector. and standard deviation These two vectors together parameterize a Gaussian distribution. This is the input signal. Potential representations in the potential space.

[0024] In this embodiment, the generator is the core of the SoloRadar generative model. It abandons the traditional approach of separating interference from mixed signals and instead, based on the latent representation of the cue signal in the latent space constructed by the VAE, it... Potential representation for directly generating pure target echoes To achieve this goal, a diffusion transformer (uDiT) based on the U-Net architecture is used as the generator. By combining the classic U-Net skeleton with the powerful Transformer computing unit and introducing a special guidance mechanism, the predicted velocity field is output, which indicates the direction in which to move in order to efficiently return to a clean image from the current noisy state. This predicted velocity field is used to efficiently purify the time-frequency characteristics of the radar signal and obtain a potential representation of the clean target echo.

[0025] Please see Figure 4 , Figure 4 This is a schematic diagram of the diffusion converter based on the U-Net architecture provided in an embodiment of the present invention, as shown below. Figure 4As shown, the diffusion transformer based on the U-Net architecture in this embodiment includes a time feature extraction module, a conditional feature extraction module, and a purification module. The time feature extraction module includes a connected encoding module and an adaptive normalization module. The diffusion time step is encoded by the encoding module and then input into the adaptive normalization module to generate time features. The conditional feature extraction module includes a connected first MLP and a guiding transformer. The latent representation of the cue signal is passed through the first MLP and then input into the guiding transformer to extract conditional features. The purification module includes a connected second MLP, a purification transformer, and a third MLP. The latent representation of the echo signal is passed through the second MLP and then input into the purification transformer. The purification transformer performs feature purification and information fusion on the input vector based on the conditional and time features, and then passes it through the third MLP to obtain the target latent representation.

[0026] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of the guide converter provided in an embodiment of the present invention, as shown below. Figure 5 As shown, the guiding transformer includes an input linear projection layer, a Transformer encoder, a Norm layer, and an output linear projection layer connected in sequence. A position encoding layer is connected between the input end of the input linear projection layer and the input end of the Transformer encoder.

[0027] In this embodiment, the role of the guide converter is to extract the potential representation of the cue signal. In this process, a set of high-level conditional features that can accurately describe the intrinsic structure of the target signal are extracted. For example... Figure 5 As shown, the latent features of the input are first mapped to the high-dimensional embedding space of the model through an input linear projection layer, and positional encoding is added to them. Subsequently, the sequence is fed into a standard multi-layer Transformer encoder, which, through a multi-head self-attention mechanism, can capture the complex temporal and frequency dependencies within the reference waveform, and finally outputs a set of refined conditional features that can serve as waveform fingerprints.

[0028] Please continue reading Figure 4 The purification converter comprises multiple uDiT modules arranged in a symmetrical U-shape, forming encoding and decoding paths. The encoding and decoding paths are connected across the same levels. The specific structure of the uDiT module is as follows: Figure 6 The diagram shown is a structural schematic of the uDiT module provided in this embodiment of the invention. The uDiT module includes a first Layer Norm layer, a first Scale Shift layer, a multi-head self-attention block, a first Scale layer, a multi-head cross-attention block, a second Layer Norm layer, a second Scale Shift layer, a second Scale layer, a fourth MLP, a fifth MLP, and a sixth MLP.

[0029] The system consists of a first Layer Norm layer, a first Scale Shift layer, a multi-head self-attention block, and a first Scale layer connected sequentially. The output of the first Scale layer is added to the input residual of the first Layer Norm layer and then input to the multi-head cross-attention block. The input and output residuals of the multi-head cross-attention block are added to the second Layer Norm layer. The second Layer Norm layer, the second Scale Shift layer, the fourth MLP, and the second Scale layer are connected sequentially. The output of the second Scale layer is added to the input residual of the second Layer Norm layer and then used as the output of the uDiT module. Conditional features are input to the multi-head cross-attention block after passing through the fifth MLP. Temporal features are generated by the sixth MLP into a first scaling parameter γ1, a first translation parameter β1, a first scaling factor α1, a second scaling parameter γ2, a second translation parameter β2, and a second scaling factor α2. The first scaling parameter γ1 and the first translation parameter β1 are input to the first Scale Shift layer, the first scaling factor α1 is input to the first Scale layer, the second scaling parameter γ2 and the second translation parameter β2 are input to the second Scale Shift layer, and the second scaling factor α2 is input to the second Scale layer.

[0030] In this embodiment, the uDiT module is the core computing unit of the refinement converter, responsible for performing specific feature refinement and information fusion tasks. Internally, the data stream first captures the internal structure of the signal itself through multi-head self-attention (MHSA). Subsequently, in the most critical multi-head cross-attention (MHCA) step, the model uses the main path signal as a query, actively referencing and aligning the conditional features that guide the converter output, thereby achieving accurate differentiation between the target and interference.

[0031] It should be noted that the behavior of the purification transformer is precisely controlled by the diffusion time step t. The diffusion time step encoding not only dynamically generates scaling (γ) and translation (β) parameters to adjust the statistical scale of the signal through time adaptive normalization (AdaLN), but also generates a scaling factor (α) through a gated residual connection mechanism to flexibly control the fusion weights of newly extracted features. This dual adjustment mechanism ensures that the purification process remains stable and efficient under different noise levels.

[0032] Following the reverse process of the diffusion model, the generator progressively reconstructs the latent representation of the target from random Gaussian noise, generating an estimated latent representation of the target. Then, a preliminary reconstructed signal is obtained from the estimated latent representation of the target using a decoder based on a variational autoencoder. Please continue reading. Figure 3The decoder, in contrast to the encoder, is tasked with reconstructing vectors in the latent space back to the original time-domain signal. The decoder receives the target latent representation and, through a series of TF-GridNet modules and transposed convolutional layers, progressively upsamples it to recover the dimensions and structure of the time-frequency map. Finally, through inverse Fourier transform, the reconstructed complex time-frequency map is converted back to a time-domain I / Q signal, i.e., the initial reconstructed signal.

[0033] Understandably, the initial reconstructed signal obtained after passing through the generator and VAE decoder, while recovering the main structure of the target echo, may still contain computational artifacts or phase distortion due to inherent losses in the VAE reconstruction process and potential minor biases in the generator. To further improve signal fidelity, especially for radar applications with extremely high coherence requirements, a corrector based on a fractional single-step diffusion model is employed to perform final fine adjustments to the initial reconstructed signal.

[0034] In this embodiment, the fractional single-step diffusion model employs the NCSN++ backbone network and operates directly in the time-spectrum domain. Its purification process is guided by two key conditions: one is the echo signal, i.e., the target echo signal received by the radar containing interference and noise signals, which provides complete contextual information; the other is the preliminary reconstruction signal from the previous stage, serving as a guide for high-quality repair.

[0035] It should be noted that the SoloRadar generative model in this embodiment of the invention can be trained in stages.

[0036] The first stage involves training the variational autoencoder separately, during which the generator and corrector are not involved in the training. The training data includes clean radar echo signals and target echo signals containing interference as the reconstructed targets. The loss function includes reconstruction loss and KL divergence loss.

[0037] The second stage involves training the generator separately. In this stage, the corrector is not involved in the training, and the parameters of the variational autoencoder are frozen. A pre-trained VAE encoder is used to convert the time-domain signal into a latent representation, treating it as a fixed, invariant data source. The echo signal, cue signal, and clean target echo signal are each passed through the frozen VAE encoder to obtain their latent representations, which are then used as training data. During training, noise is progressively added to the latent representation of the clean target echo signal. The loss function can be designed to minimize the difference between the velocity field predicted by the uDiT network and the actual velocity field. This teaches the generator how to utilize the conditions of the cue signal to progressively denoise and reconstruct the latent representation of the clean target echo signal in the latent space.

[0038] In the third stage, the corrector is trained separately. The parameters of the variational autoencoder and generator are frozen. The echo signal is passed through the VAE encoder, generator, and VAE decoder to obtain a preliminary reconstructed signal. A random masking enhancement strategy is applied to the preliminary reconstructed signal, randomly discarding a portion of information to simulate potential flaws and incompleteness produced by the generator. This forces the corrector to search for clues in the original echo signal for repair, thereby learning a more robust purification capability. Using the scale-invariant signal-to-noise ratio (SI-SNR) in the signal domain as the loss function ensures that the final output waveform has maximum fidelity and coherence.

[0039] The radar anti-jamming method based on the cascaded generative model of the present invention adopts a generative purification method, which eliminates the need to establish an accurate mathematical model of the interference. Through a diffusion model guided by the cue signal, it can directly generate a pure target echo signal from the potential space. After the signal is purified in the potential space, fine repair is performed, achieving high-fidelity signal recovery and solving the signal distortion problem caused by discriminative deep learning methods.

[0040] Furthermore, simulation experiments are used to illustrate the effectiveness of the radar anti-jamming method based on the cascaded generative model of this invention.

[0041] 1. Dataset To better reflect real-world interference scenarios, we set some key parameters to be randomly generated within a certain range. The specific parameters are shown in Table 1. A total of 100 coherent processing intervals were generated, of which 75% were used for training, 15% for validation, and 10% for testing.

[0042] Table 1

[0043] 2. Evaluation Indicators Various anti-jamming measures have varying effects on suppressing different types of interference. Accurate qualitative and quantitative evaluation of these measures is crucial for establishing intelligent anti-jamming models for radar. In radar signal processing systems, the ultimate goal of any anti-jamming measure is to reduce the power of the interfering signal and detect targets more effectively. Therefore, the signal-to-interference-plus-noise ratio (SNR) after implementing these measures is used as a performance indicator to measure the effectiveness of the anti-jamming measures, defined as the interference suppression ratio. : ; in, This represents the signal-to-interference-plus-noise ratio (SIR) of the signal before anti-interference measures are implemented. This indicates the signal-to-interference-plus-noise ratio (SIR) of the signal after anti-interference measures have been implemented. Indicates the target signal. Indicates a recovery signal. This is the received echo containing interference. Calculation. To obtain in decibels Value. Based on The concept can be used to establish interference suppression evaluation criteria: The larger the value, the better the interference suppression effect.

[0044] 3. Simulation Results and Analysis 3.1 Visual Analysis of Interference Suppression Effect To visually demonstrate the interference suppression effect of the proposed SoloRadar generative model, a set of typical simulation data was selected, with an input signal-to-noise ratio of 5dB and an interference-to-signal ratio of 20dB. The time-domain and time-frequency waveforms of the signal before and after processing are shown below. Figure 7 and Figure 8 As shown, Figure 7 These are time-domain waveform comparison diagrams of the simulation experiments provided in this embodiment of the invention; Figure 8 This is a comparison diagram of the time-frequency domain waveforms of the simulation experiment provided in the embodiment of the present invention.

[0045] Comparative analysis of the time and time-frequency domains demonstrates that the SoloRadar generative model not only recovers the macroscopic envelope of the signal in the time domain but also accurately identifies and eliminates ISRJ interference with similar characteristics from a detailed two-dimensional time-frequency structure, while maximizing the preservation of the time-frequency integrity of the real target signal. This intuitively and powerfully proves the effectiveness and precise suppression capability of our proposed algorithm.

[0046] 3.2 Analysis of Anti-interference Effect Time-varying filtering algorithms are now relatively mature and widely used in signal interference suppression. Among them, the STFT and STFRFT time-varying filtering algorithms suppress interference signals by reconstructing the target signal, and have achieved certain results in the field of radar interference suppression. Therefore, the SoloRadar generative model proposed in this paper is compared with the STFT and STFRFT time-varying filtering algorithms with fixed and adaptive window lengths, respectively. The average interference suppression ratio results of each method are shown in Table 2.

[0047] Table 2

[0048] The data shows that the SoloRadar generative model significantly outperforms other methods in interference suppression. The high performance of the SoloRadar generative model is attributed to its ability to learn the basic characteristics of the target signal and improve network performance through continuous iterative optimization of the diffusion model. Therefore, it can recover the target signal from echo signals carrying interference and noise, and its interference suppression performance is superior to existing methods.

[0049] 3.3 Analysis of Parameter Measurement Results The performance of an advanced anti-jamming algorithm is evaluated not only by its effectiveness in suppressing interference, but also by its ability to faithfully preserve key target parameters. Therefore, Monte Carlo simulations are used to quantitatively assess the accuracy of distance and velocity estimation of the target after processing by the SoloRadar generative model.

[0050] Accurate estimation of target range and velocity relies on pulse compression and coherent accumulation processing, respectively. One of the core advantages of the proposed SoloRadar generative model lies in its end-to-end complex network architecture. This architecture can directly process the I / Q complex data of the radar echo, thereby suppressing interference while preserving the phase information, which is crucial for parameter measurement, to the maximum extent.

[0051] Please see Figure 9 and Figure 10 The comparison results of pulse compression and coherent accumulation (Doppler spectrum) of the echo signal before and after processing by the method of this invention are shown. Figure 9 This is a comparison chart of pulse compression results before and after interference processing provided in an embodiment of the present invention; Figure 10 This is a comparison diagram of the coherent results before and after interference processing provided in the embodiments of the present invention.

[0052] As can be seen from the figure, the unprocessed signal exhibits blurred pulse compression peaks and a chaotic Doppler spectrum, making effective parameter extraction impossible. After processing with the SoloRadar generative model, the signal pulse compression results show sharp and clear peaks, and coherent accumulation forms a significant energy concentration on the target's Doppler cells. This intuitively demonstrates that the SoloRadar generative model not only restores target detectability but also lays a solid foundation for subsequent high-precision parameter estimation.

[0053] To quantitatively evaluate the parameter estimation accuracy of the algorithm, we conducted 1000 independent Monte Carlo simulations and calculated the root mean square error (RMSE) of distance and velocity measurements. This metric accurately reflects the average fluctuation between the estimated and true values. The parameter measurement accuracy is shown in Table 3. After processing using the method of this invention, the RMSEs for distance and velocity estimation are only 0.76 m and 0.28 m / s, respectively. This further demonstrates the advantages of the method of this invention.

[0054] Table 3

[0055] In a second aspect, embodiments of the present invention provide a radar anti-jamming device based on a cascaded generative model, applicable to the radar anti-jamming method based on a cascaded generative model in the first aspect. The device includes: The signal acquisition unit is used to acquire echo signals and prompt signals; The generation unit is used to input the echo signal and cue signal into the trained SoloRadar generative model to generate a clean target echo signal; The SoloRadar generative model compresses the echo signal and cue signal into the latent space to obtain the corresponding latent representation. The latent representation of the cue signal is used as a guiding condition to denoise and extract the latent representation of the echo signal to obtain the target latent representation. After decoding and correcting the target latent representation, a clean target echo signal is obtained.

[0056] For details regarding the radar anti-jamming device based on the cascaded generative model and its corresponding beneficial effects, please refer to the relevant content of the radar anti-jamming method based on the cascaded generative model provided in the first aspect, which will not be repeated here.

[0057] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0058] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0059] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A radar anti-jamming method based on a cascaded generative model, characterized in that, include: Step 1: Acquire echo signals and prompt signals; Step 2: Input the echo signal and the prompt signal into the trained SoloRadar generative model to generate a clean target echo signal; The SoloRadar generative model compresses the echo signal and the cue signal into the latent space to obtain corresponding latent representations. Using the latent representation of the cue signal as a guiding condition, it denoises and extracts the latent representation of the echo signal to obtain the target latent representation. After decoding and correcting the target latent representation, the clean target echo signal is obtained.

2. The radar anti-jamming method based on a cascaded generative model according to claim 1, characterized in that, The echo signal is a target echo signal received by the radar that contains interference and noise signals, and the cue signal is a priori signal containing target characteristics.

3. The radar anti-jamming method based on a cascaded generative model according to claim 1, characterized in that, The SoloRadar generative model includes: a variational autoencoder operating in the time-frequency domain, a generator, and a corrector; the variational autoencoder includes an encoder and a decoder; wherein... The encoder is used to perform short-time Fourier transform on the input echo signal and the prompt signal to obtain the corresponding spectrogram, and the TF-GridNet module is used to compress the spectrogram into the latent space to obtain the corresponding latent representation; The generator is a diffusion transformer based on the U-Net architecture, used to denoise and extract the latent representation of the echo signal according to the guiding conditions, so as to obtain the target latent representation, which is the latent representation of the clean target echo; The encoder is used to reconstruct the target latent representation into a corresponding time-frequency map using the TF-GridNet module, and to recover the time-frequency map into a preliminary reconstructed signal through inverse Fourier transform; The corrector is a fractional single-step diffusion model used to obtain a clean target echo signal based on the preliminary reconstructed signal and the echo signal.

4. The radar anti-jamming method based on a cascaded generative model according to claim 3, characterized in that, The diffusion transformer based on the U-Net architecture includes: a temporal feature extraction module, a conditional feature extraction module, and a purification module, wherein, The time feature extraction module includes a connected encoding module and an adaptive normalization module. The diffusion time step is encoded by the encoding module and then input into the adaptive normalization module to generate time features. The conditional feature extraction module includes a first MLP and a guided transformer connected together. The latent representation of the cue signal is extracted by the guided transformer after being passed through the first MLP to obtain conditional features. The purification module includes a second MLP, a purification transformer, and a third MLP connected together. The latent representation of the echo signal is input to the purification transformer after passing through the second MLP. The purification transformer performs feature purification and information fusion on the input vector according to the conditional features and the time features, and then obtains the target latent representation through the third MLP.

5. The radar anti-jamming method based on a cascaded generative model according to claim 4, characterized in that, The guided transformer includes an input linear projection layer, a Transformer encoder, a Norm layer, and an output linear projection layer connected in sequence, with a position encoding layer connected between the input end of the input linear projection layer and the input end of the Transformer encoder.

6. The radar anti-jamming method based on a cascaded generative model according to claim 4, characterized in that, The purification converter includes multiple uDiT modules arranged in a symmetrical U-shaped structure to form an encoding path and a decoding path, with skip connections between the same levels in the encoding path and the decoding path.

7. The radar anti-jamming method based on a cascaded generative model according to claim 6, characterized in that, The uDiT module includes a first Layer Norm layer, a first Scale Shift layer, a multi-head self-attention block, a first Scale layer, a multi-head cross-attention block, a second Layer Norm layer, a second Scale Shift layer, a second Scale layer, a fourth MLP, a fifth MLP, and a sixth MLP; wherein, The first Layer Norm layer, the first Scale Shift layer, the multi-head self-attention block, and the first Scale layer are connected in sequence; the output of the first Scale layer is added to the input residual of the first Layer Norm layer and then input to the multi-head cross-attention block; the input and output residuals of the multi-head cross-attention block are added to the second Layer Norm layer; the second Layer Norm layer, the second Scale Shift layer, the fourth MLP, and the second Scale layer are connected in sequence; the output of the second Scale layer is added to the input residual of the second Layer Norm layer and then used as the output of the uDiT module; The conditional features are then input to the multi-head cross-attention block after passing through the fifth MLP; The time features are generated by the sixth MLP into a first scaling parameter, a first translation parameter, a first scaling factor, a second scaling parameter, a second translation parameter, and a second scaling factor; the first scaling parameter and the first translation parameter are input to the first Scale Shift layer, the first scaling factor is input to the first Scale layer, the second scaling parameter and the second translation parameter are input to the second Scale Shift layer, and the second scaling factor is input to the second Scale layer.

8. The radar anti-jamming method based on a cascaded generative model according to claim 3, characterized in that, The fraction-based single-step diffusion model employs the NCSN++ backbone network.

9. A radar anti-jamming device based on a cascaded generative model, characterized in that, The radar anti-jamming method based on the cascaded generative model as described in any one of claims 1-8 includes: The signal acquisition unit is used to acquire echo signals and prompt signals; The generation unit is used to input the echo signal and the prompt signal into the trained SoloRadar generative model to generate a clean target echo signal; The SoloRadar generative model compresses the echo signal and the cue signal into the latent space to obtain corresponding latent representations. Using the latent representation of the cue signal as a guiding condition, it denoises and extracts the latent representation of the echo signal to obtain the target latent representation. After decoding and correcting the target latent representation, the clean target echo signal is obtained.

Images