Radar active deception jamming signal recognition method based on denoising diffusion probability model
The radar active deception jamming signal identification method based on the denoising diffusion probability model uses the tag-guided denoising diffusion probability model to generate signal data with the same distribution as the real signal, which solves the problem of insufficient active deception jamming samples and improves the identification accuracy and model training efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ANHUI UNIV
- Filing Date
- 2023-09-07
- Publication Date
- 2026-06-23
AI Technical Summary
Existing technologies struggle to capture a sufficient number of active deception interference samples within a limited timeframe, leading to a decline in the recognition performance of feature extraction machine learning and deep learning recognition methods. Generative adversarial networks cannot capture the full diversity of target data distribution, resulting in limited similar samples that fail to meet the requirements for sample diversity.
A label-guided denoising diffusion probability model is established by acquiring original time-frequency image samples through short-time Fourier transform based on a denoising diffusion probability model. The model is trained using a U-net neural network to generate signal data with the same distribution as the real signal. An active deception interference signal dataset is constructed and divided into training, validation and test sets. An interference identification model is then established for identification.
It improves the recognition accuracy of radar active deception jamming signals, the generated data can effectively expand the training set, improve the performance of the recognition model, overcome the problems of intra-class similarity and inter-class difference of samples in low signal-to-noise ratio scenarios, the generated samples are of high quality, and the recognition model converges quickly.
Smart Images
Figure CN117368857B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radar interference identification technology, and in particular to a method for identifying active deception interference signals in radar based on a denoising diffusion probability model. Background Technology
[0002] For radar, identifying active deception jamming from the acquired echoes and implementing effective jamming suppression measures is crucial. However, the dynamic and rapidly changing electromagnetic environment of the modern battlefield poses a significant challenge to radar systems, especially when facing adversarial active jammers within the limited operational timeframe, requiring the acquisition of a sufficient number of deception jamming samples. Therefore, a lack of jamming samples will inevitably and severely degrade the recognition performance of feature extraction machine learning methods and deep learning recognition methods.
[0003] Given the importance of the richness and diversity of training data, researchers have employed numerous data augmentation methods to increase the dimensionality of training samples. Common manual data augmentation methods, such as rotation, cropping, and stitching, are inefficient and, to some extent, fail to meet the requirements for sample diversity. In particular, Generative Adversarial Networks (GANs) are widely used for data augmentation due to their ability to recover the original data distribution and have been successfully applied to generate synthetic aperture radar (SAR) images. The Wasserstein GAN with gradient penalty (WGAN-GP) augmented SAR image datasets to approximate real images and improved the recognition efficiency of Support Vector Machines (SVMs) for target identification with limited samples. Despite their advantages in data augmentation, it is worth noting that GANs often suffer from mode collapse because they fail to capture the full diversity of target data distributions, instead producing a limited set of similar samples lacking variation or representing only a subset of the desired patterns. Summary of the Invention
[0004] To address the technical problems existing in the background art, this invention proposes a radar active deception jamming signal identification method based on a denoising diffusion probability model.
[0005] This invention proposes a method for identifying active radar deception jamming signals based on a denoising diffusion probability model, comprising:
[0006] Acquire radar active deception jamming signals with jamming type labels;
[0007] Short-time Fourier transform was used to perform time-frequency analysis on the radar active deception jamming signal to obtain the original sample set of time-frequency images;
[0008] Establish a label-guided denoising diffusion probability model;
[0009] The label-guided denoising and diffusion probability model is trained based on the original sample set of time-frequency images to obtain the trained label-guided denoising and diffusion probability model.
[0010] Based on the trained label-guided denoising diffusion probability model and interference type labels, time-frequency image generation samples are generated, and a time-frequency image generation sample set is established based on the time-frequency image generation samples.
[0011] A dataset of active deception interference signals is established based on the original sample set of time-frequency images and the sample set generated from time-frequency images.
[0012] The active spoofing jamming signal dataset is divided into training, validation, and test sets;
[0013] Establish an interference identification model;
[0014] The interference recognition model was trained, validated, and tested sequentially using the training set, validation set, and test set, and the interference recognition model that passed the test was obtained.
[0015] The tested and approved interference identification model was used to identify active deception jamming signals on radar.
[0016] Preferably, label
[0017] The guided denoising diffusion probability model includes the U-net neural network; the U-net network structure includes 14 residual blocks, 2 convolutional layers, 2 downsampling layers, 2 upsampling layers and one attention block; among them, the residual module introduces a noise step and an interference type label embedding vector.
[0018] Preferably, the label-guided denoising diffusion probability model is trained using the training dataset to obtain a trained label-guided denoising diffusion probability model, specifically including:
[0019] A time-frequency image sample (x0, y) is randomly selected from the original time-frequency image sample set; where x0 represents the image at the noise addition step t = 0, and y represents the interference type label of x0;
[0020] Randomly sample a t from {1, 2, ..., T}; where t represents the current noise-adding step number and T represents the total noise-adding step number.
[0021] Randomly generated noise In the formula, ∈ t This represents Gaussian noise that follows a standard normal distribution.
[0022] Based on randomly generated noise and the original time-frequency image samples, calculate the noise image x obtained by adding noise. t ;in, In the formula, x tThis represents the image obtained after adding noise for t steps, where x0 represents the image at the point where the number of noise addition steps t = 0. β e Indicates the noise variance;
[0023] The noisy image, interference type label, and noise addition step count are input into the label-guided denoising diffusion probability model to obtain the predicted noise ∈ θ (x t ,y,t);
[0024] Calculate the loss function L based on the randomly generated noise and the predicted noise;
[0025] Where, L=||∈ t -∈ θ (x t ,y,t)|| 2 In the formula, L represents the loss function, and |||| represents the L2 norm;
[0026] The parameters θ of the label-guided denoising diffusion probability model are iteratively updated using the backpropagation algorithm until the loss function converges, thus obtaining the trained label-guided denoising diffusion probability model.
[0027] Preferably, based on the trained label-guided denoising diffusion probability model and interference type labels, time-frequency image generation samples are generated, specifically including:
[0028] Obtain the trained label-guided denoising diffusion probability model to determine the generated subsequence {τ0, τ1, τ2, ..., τ}. S} and the interference type label y to be generated; where τ0=0, τ S =T;
[0029] Randomly sample noisy images
[0030] Based on the noise image The noise image of the previous noise-adding step is calculated.
[0031] in,
[0032] In the formula, Indicates generation after τ s The generated image obtained by adding noise in one step. Indicates generation after τ s-1 The generated image obtained by adding noise in one step. express The corresponding prediction noise, This refers to the pre-defined hyperparameters. This represents noise obtained from random sampling. This means t = τ Sα at time t value, This means t = τ S-1 α at time t value, β e Indicates the noise variance;
[0033] Iteratively calculate the noise image of the previous noise-adding step in the reverse process until τ0 = 0, generating time-frequency image samples.
[0034] Preferably, the jamming type labels include: range spoofing jamming labels, velocity spoofing jamming labels, dense false target jamming labels, intermittent sampling direct forwarding jamming labels, intermittent sampling repeated forwarding jamming labels, intermittent sampling cyclic forwarding jamming labels, mixed labels of range spoofing jamming and intermittent sampling repeated forwarding jamming, mixed labels of range spoofing jamming and intermittent sampling direct forwarding jamming, mixed labels of velocity spoofing jamming and intermittent sampling direct forwarding jamming, and mixed labels of range spoofing jamming and velocity spoofing jamming.
[0035] Preferably, the active spoofing jamming signal dataset is divided into a training set, a validation set, and a test set, specifically including:
[0036] The active deception jamming signal dataset is divided into corresponding data subsets according to the jamming type label;
[0037] Samples are randomly selected from each data subset according to a set ratio to form the training set, validation set, and test set.
[0038] Preferably, the ratio of the training set, validation set, and test set is 7:2:1.
[0039] Preferably, the interference identification model is an AlexNet model, a support vector machine model, a decision tree model, or a logistic regression model.
[0040] The present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the radar active deception jamming signal identification method based on the denoising diffusion probability model as described in any of the above.
[0041] The present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the radar active deception jamming signal identification method based on the denoising diffusion probability model as described in any of the above.
[0042] This invention proposes a radar active deception jamming signal identification method based on a denoising diffusion probability model. It constructs a dataset of active deception jamming signals received by radar under active jamming conditions. During the construction process, short-time Fourier transform (STFT) is used to perform time-frequency analysis on the acquired small number of radar active deception jamming signals, obtaining a raw time-frequency image sample set for a more comprehensive understanding and visualization of the signal's time-frequency characteristics. The established label-guided denoising diffusion probability model is then trained based on this raw time-frequency image sample set, resulting in a well-trained model. This well-trained model is then used to quickly generate a large amount of signal data with the same distribution as the real signal. This expanded data effectively improves the accuracy of the subsequent trained jamming identification model in identifying radar active deception jamming signals. Attached Figure Description
[0043] Figure 1 This is a flowchart illustrating a radar active deception jamming signal identification method based on a denoising diffusion probability model, as proposed in one embodiment of the present invention.
[0044] Figure 2 This is a schematic diagram of the forward and reverse processes of the tag-guided denoising diffusion probability model in one embodiment of the present invention.
[0045] Figure 3 This is a schematic diagram of the structure of the U-net neural network in one embodiment of the present invention.
[0046] Figure 4 In one embodiment of the present invention, five samples are randomly selected from each interference category label in a -7dB scene.
[0047] Figure 5 This is a generated sample based on five samples when using the traditional WGAN-GP method in a -7dB scenario, as proposed in one embodiment of the present invention.
[0048] Figure 6 This is a generated sample based on five samples when using the method described in this invention in a -7dB scenario, according to one embodiment of the present invention.
[0049] Figure 7 The classification accuracy is calculated by comparing the traditional WGAN-GP method and the method described in this invention when adding different numbers of generated samples in a -5dB scenario.
[0050] Figure 8 The classification accuracy is calculated by comparing the traditional WGAN-GP method and the method described in this invention when adding different numbers of generated samples in a -6dB scenario.
[0051] Figure 9The classification accuracy is calculated by comparing the traditional WGAN-GP method and the method described in this invention when adding different numbers of generated samples in a -7dB scenario.
[0052] Figure 10 The classification accuracy is calculated by comparing the traditional WGAN-GP method and the method described in this invention when adding different numbers of generated samples in a -8dB scenario.
[0053] Figure 11 The classification accuracy is calculated by comparing the traditional WGAN-GP method and the method described in this invention when adding different numbers of generated samples in a -9dB scenario.
[0054] Figure 12 The classification accuracy is calculated by comparing the traditional WGAN-GP method and the method described in this invention when adding different numbers of generated samples in a -10dB scenario. Detailed Implementation
[0055] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0056] Reference Figure 1 The present invention proposes a method for identifying radar active deception jamming signals based on a denoising diffusion probability model, comprising:
[0057] Construct a dataset of active deception jamming signals received by radar under active jamming conditions;
[0058] The active spoofing jamming signal dataset is divided into training, validation, and test sets;
[0059] Establish an interference identification model;
[0060] The interference recognition model was trained, validated, and tested sequentially using the training set, validation set, and test set, and the interference recognition model that passed the test was obtained.
[0061] The tested and approved interference identification model was used to identify active radar deception jamming signals.
[0062] Specifically, the construction of a dataset of active deception jamming signals received by radar under active jamming conditions includes:
[0063] Acquire radar active deception jamming signals with jamming type labels;
[0064] Short-time Fourier transform (STFT) is used to perform time-frequency analysis on radar active deception jamming signals to obtain the original sample set of time-frequency images;
[0065] Establish a label-guided denoising diffusion probability model (LG-DDPM);
[0066] The LG-DDPM model is trained based on the original sample set of time-frequency images to obtain the trained LG-DDPM model;
[0067] Input the original time-frequency image sample set into the trained LG-DDPM model to establish the time-frequency image generation sample set;
[0068] An active deception interference signal dataset is constructed based on the original sample set of time-frequency images and the sample set generated from time-frequency images.
[0069] To obtain high-quality generated images, this invention constructs a dataset of active deception jamming signals received by radar under active jamming conditions. During the construction process, Short Time Fourier Transform (STFT) is used to perform time-frequency analysis on the acquired small amount of radar active deception jamming signals to obtain a raw time-frequency image sample set, thereby gaining a more comprehensive understanding and visualization of the signal's time-frequency characteristics. Based on the raw time-frequency image sample set, the established LG-DDPM model is trained to obtain a trained LG-DDPM model. Then, the trained LG-DDPM model is used to quickly generate a large amount of signal data with the same distribution as the real signal. The expanded data can effectively improve the accuracy of the jamming identification model in identifying radar active deception jamming signals.
[0070] In this embodiment, the LG-DDPM model includes a forward process (adding noise) and a reverse process (denoising), such as... Figure 2 As shown.
[0071] In the forward process, the data xt obtained after adding noise to the original data x0~q(x) for t steps can be directly calculated using the parameter renormalization formula:
[0072]
[0073] Where, x T Let x represent the image data after T rounds of noise addition. T It conforms to a standard Gaussian distribution; T represents the total number of noise addition steps, t represents the current number of noise addition steps, and x0 represents the image data at the time of t=0 noise addition steps. α t It can be derived from the variance scheme {β1, β2, ..., β...} T} Determine; where β t This represents the noise variance when the number of noise-adding steps in the forward process is t; ∈ t This represents Gaussian noise with a mean of 0 and a variance of 1, denoted as .
[0074] The reverse process involves generating target data using standard Gaussian noise. This requires a trained network model to predict the noise introduced during the forward process; this network is called the noise prediction network model, denoted as ∈ θ (x t (x, y, t), where y represents the interference type label of data x0. Specifically, x t-1 Can be derived from x t Based on formula (2), the following can be derived:
[0075]
[0076] In equation (2), σ t This represents the pre-defined hyperparameters.
[0077] The parameters θ in the noise prediction network model are optimized using the loss function of equation (3):
[0078]
[0079] in, This represents the loss function at step t-1. This indicates that the mean of a function is calculated based on the variables; |||| represents the L2 norm.
[0080] It is worth noting that all σ can share the same noise prediction network model. In other words, when σ changes, there is no need to retrain the noise prediction network model.
[0081] when When the reverse process conforms to the Markov decision process, the noise prediction network model is the LG-DDPM model.
[0082] Considering that the denoising objective function is independent of a specific reverse process, training can be attempted on a forward process of length less than T, which can speed up the generation process without training multiple models. Therefore, the proposed LG-DDPM uses an increasing subsequence τ = {τ1, τ2, ..., τ...} of length S < T. S Therefore, the reverse process is no longer caused by x. t Generate x t-1 , but by generate The generation process can be rewritten as Equation (4). Ultimately, this reverse process does not require calculating the noisy image X at each noise-adding step. t This accelerates the reverse process.
[0083] Equation (4) is:
[0084]
[0085] in, Indicates generation after τs The generated image obtained by adding noise in one step. It represents the generation after τ s-1 The generated image obtained by adding noise in one step. express The corresponding prediction noise, This means t = τ S α at time t value, This means t = τ S α at time t value, This refers to the pre-defined hyperparameters. This represents noise obtained from random sampling.
[0086] To facilitate the integration of interference type labels and time with image components, such as... Figure 3 As shown, in this embodiment, the label-guided denoising diffusion probability model, also known as the noise prediction network model, includes the U-net neural network. The U-net network structure includes 14 residual blocks, 2 convolutional layers, 2 downsampling layers, 2 upsampling layers, and one attention block. The residual blocks incorporate noise-adding steps and interference type label embedding vectors.
[0087] In particular, in order to integrate label information into the U-net stream, the residual module in this embodiment uses cross-attention blocks to asymmetrically combine the embedded sequences of labels and images, with labels as input to the query and images as input to keywords and values, enabling U-net to focus on label-related information.
[0088] During the training phase, the noise step and interference type labels are encoded by the embedding layer and then merged with the residual block as input to the network. The output of U-net is the predicted noise, which is combined with sampled Gaussian noise to calculate the loss, and then the parameters of U-net are updated through backpropagation. During the generation phase, the output of U-net is close to Gaussian noise and is used to derive the image data of the previous noise step number during the backpropagation process according to Equation (1). The difference is that the training phase always requires the noise step number t, but the generation phase does not require the continuous derivation of the data of the previous noise step number.
[0089] In this embodiment, the label-guided denoising diffusion probability model is trained using the training dataset to obtain a trained label-guided denoising diffusion probability model, specifically including:
[0090] One original time-frequency image sample is randomly selected from the original time-frequency image sample set; wherein, the original time-frequency image sample set is... The original samples of the time-frequency image are (x0, y), and the variance scheme is {β1, β2, ..., β...}. TIn the dataset, N represents the total number of original time-frequency image samples in the dataset, and n represents the number of a specific original time-frequency image sample in the dataset.
[0091] Randomly sample a value t from {1, 2, ..., T};
[0092] Randomly generated noise In the formula, ∈ t This represents Gaussian noise that follows a standard normal distribution.
[0093] Calculate the image x obtained by adding noise T ;in, In the formula, x t This represents the image obtained after adding noise for t steps, where x0 represents the image at t=0. β e This represents the noise variance of the forward process at step t.
[0094] Image x t The interference type label y and the number of noise addition steps t are used as input labels to guide the denoising diffusion probability model, resulting in the predicted noise ∈ θ (x t ,y,t);
[0095] Calculate the loss function L based on the randomly generated noise and the predicted noise; where L=||∈ t -∈ θ (x t ,y,t)|| 2 ;
[0096] The parameters θ of the label-guided denoising diffusion probability model are iteratively updated using the backpropagation algorithm until the loss function converges, thus obtaining the trained label-guided denoising diffusion probability model.
[0097] To quickly generate time-frequency image generation samples and thus rapidly establish a time-frequency image generation sample set to improve the recognition accuracy of the interference identification model, in this embodiment, time-frequency image generation samples are generated based on a trained label-guided denoising diffusion probability model and interference type labels. Specifically, this includes:
[0098] Obtain the trained label-guided denoising diffusion probability model to determine the generated subsequence {τ0, τ1, τ2, ..., τ}. S} and the interference type label y to be generated; where τ0=0, τ S =T;
[0099] Randomly sample noisy images in,
[0100] Based on the noise image Calculate the noise image of the previous noise addition step.
[0101] in,
[0102] in, Indicates passing through τ s The image obtained by adding noise in one step, It indicates that after τ s-1 The image obtained by adding noise in one step, express The corresponding prediction noise, This means t = τ S α at time t value, This means t = τ S α at time t value, This refers to the pre-defined hyperparameters. Indicates to The noise obtained from the actual sampling;
[0103] Iteratively calculate the noise image from the previous noise-adding step until τ0 = 0, generating time-frequency image samples.
[0104]
[0105] In this embodiment, acquiring radar active deception jamming signals with jamming type labels specifically includes:
[0106] Active deception jamming signals with jamming type labels are obtained by performing operations such as delaying and modulating radar transmitted signals.
[0107] In this embodiment, the jamming type labels include: range spoofing jamming (DDJ) label, velocity spoofing jamming (VDJ) label, dense false target jamming (DFTJ) label, intermittent sample repeat forwarding jamming (ISRJ) label, intermittent sample direct forwarding jamming (ISDJ) label, intermittent sample cyclic forwarding jamming (ISLJ) label, mixed label of range spoofing jamming and intermittent sample repeat forwarding jamming (DDJ+ISRJ), mixed label of range spoofing jamming and intermittent sample direct forwarding jamming (DDJ+ISDJ), mixed label of velocity spoofing jamming and intermittent sample direct forwarding jamming (VDJ+ISDJ), and mixed label of range spoofing jamming and velocity spoofing jamming (DDJ+VDJ).
[0108] In this embodiment, the active spoofing interference signal dataset is divided into a training set, a validation set, and a test set, specifically including:
[0109] The active deception jamming signal dataset is divided into corresponding data subsets according to the jamming type label;
[0110] Samples are randomly selected from each data subset according to a set ratio to form the training set, validation set, and test set.
[0111] In a further embodiment, the ratio of the training set, validation set, and test set is 7:2:1.
[0112] To ensure the normal operation of subsequent models, in a further embodiment, after performing time-frequency analysis on the radar active deception jamming signal using Short Time Fourier Transform (STFT) to obtain the original time-frequency image sample set, the method further includes:
[0113] Set the size of the time-frequency image to 64×64.
[0114] In this embodiment, the interference identification model is an AlexNet model, a support vector machine model, a decision tree model, or a logistic regression model.
[0115] The present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the radar active deception jamming signal identification method based on the denoising diffusion probability model as described in any of the above claims.
[0116] The present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the radar active deception jamming signal identification method based on the denoising diffusion probability model as described in any of the preceding claims.
[0117] The effects of this invention can be further illustrated by the following simulations:
[0118] 1. Simulation conditions
[0119] The parameters of the test hardware platform are shown in Table 1, and the parameters of the software platform are shown in Table 2.
[0120] Table 1 Hardware Platform Parameters
[0121] CPU Intel i7-11700 Memory 32GB Graphics card model NVIDIA GeForce RTX 3070 Ti Graphics card memory 8GB
[0122] Table 2 Software Platform Parameters
[0123]
[0124]
[0125] 2. Simulation Method
[0126] (1) Traditional method WGAN-GP;
[0127] (2) The method described in this invention is a radar active deception jamming signal identification method based on the tag-guided denoising diffusion probability model LG-DDPM; wherein the jamming identification model is the AlexNet model.
[0128] 3. Simulation content and simulation results
[0129] To display the generated samples, five samples were randomly selected from each interference category label at -7dB, such as... Figure 4 As shown. The augmented samples generated from the five samples using the traditional WGAN-GP method are as follows: Figure 5 As shown, the augmented samples generated from five samples using the method described in this invention are as follows: Figure 6 As shown in the figure. Solid boxes indicate missing features, dashed boxes indicate incorrect features, and dotted boxes indicate blurred features.
[0130] from Figure 4 , Figure 5 and Figure 6 As shown, despite some defects in the characteristics of the generated samples, both LG-DDPM and WGAN-GP successfully generated high-quality data for each type of active spoofing interference signal in general. Both models effectively accomplished the task of generating sufficient data in an intuitive and undeniable manner. However, WGAN-GP failed to generate samples of interrupted sampling interference with a higher probability than LG-DDPM, resulting in severely impaired identification and classification of subsequent interference signals. Furthermore, some images generated by WGAN-GP exhibited a large number of noise pixels at the bottom, causing excessive blurring of features related to the interference signal, while the bottom of the LG-DDPM samples was almost solid-colored, with clear feature details of the interference signal, and did not exhibit instability.
[0131] To verify the robustness of the proposed LG-DDPM, classification was performed on the original sample set and sample sets with different numbers of generated samples added, using both the traditional method WGAN-GP and the method described in this invention, in low interference noise ratio (JNR) scenarios ranging from -5 dB to -10 dB. Classification accuracy curves were calculated and plotted for each added number of generated samples. Twenty Monte Carlo simulations were performed for each experiment to ensure the reliability of the classification results.
[0132] like Figure 7-12 As shown, from Figures 7 to 12The JNR continuously decreased, with the accuracy of classifying the original samples dropping drastically from 85.0% to 43.3% without any additional samples. During the gradual addition of samples generated by WGAN-GP and LG-DDPM to the small sample data, it was observed that although some accuracy drops or plateaus occurred at certain points, the extent of the decline was not significant. Overall, the accuracy of the samples improved significantly, especially at -10dB, where the accuracy ultimately doubled. This demonstrates the effectiveness of WGAN-GP and LG-DDPM in generating samples for small-sample classification tasks. However, in simulations where samples were added to WGAN-GP and LG-DDPM, the newly generated samples consistently showed higher classification accuracy. This indicates that the samples generated by LG-DDPM better reflect the actual data distribution, thus proving more effective in improving the overall ability of the trained model. Furthermore, during the addition of samples generated by LG-DDPM, it was found that the generated samples possessed excellent quality, and the fast convergence was highly controllable. The dynamic term added to the loss function of LG-DDPM adjusts the diversity of the sample distribution, causing the samples to increasingly reflect the actual data. More importantly, it stabilizes and simplifies the entire training process.
[0133] This application primarily studies the data augmentation problem in the task of identifying active deception interference on a small dataset. The proposed LG-DDPM overcomes the problem of its inability to adapt to intra-class similarity and inter-class differences in low JNR scenarios. Furthermore, this application establishes an expanded dataset of active deception interference samples and validates its effectiveness using AlexNet. Simulation results also demonstrate that LG-DDPM can generate higher quality samples than WGAN-GP.
[0134] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for identifying radar active deception jamming signals based on a denoising diffusion probability model, characterized in that, include: Acquire radar active deception jamming signals with jamming type labels; Short-time Fourier transform was used to perform time-frequency analysis on the radar active deception jamming signal to obtain the original sample set of time-frequency images; Establish a label-guided denoising diffusion probability model; The label-guided denoising and diffusion probability model is trained based on the original sample set of time-frequency images to obtain the trained label-guided denoising and diffusion probability model. Based on the trained label-guided denoising diffusion probability model and interference type labels, time-frequency image generation samples are generated, and a time-frequency image generation sample set is established based on the time-frequency image generation samples. A dataset of active deception interference signals is established based on the original sample set of time-frequency images and the sample set generated from time-frequency images. The active spoofing jamming signal dataset is divided into training, validation, and test sets; Establish an interference identification model; The interference recognition model was trained, validated, and tested sequentially using the training set, validation set, and test set, and the interference recognition model that passed the test was obtained. The tested and approved interference identification model was used to identify active deception jamming signals on radar.
2. The radar active deception jamming signal identification method based on the denoising diffusion probability model according to claim 1, characterized in that, The label-guided denoising diffusion probability model includes the U-net neural network; the U-net network structure includes 14 residual blocks, 2 convolutional layers, 2 downsampling layers, 2 upsampling layers, and one attention block; among them, the residual blocks introduce the number of noise addition steps and the interference type label embedding vector.
3. The radar active deception jamming signal identification method based on the denoising diffusion probability model according to claim 1, characterized in that, The label-guided denoising diffusion probability model is trained using the training dataset to obtain a well-trained label-guided denoising diffusion probability model, which specifically includes: Randomly select one original time-frequency image sample from the original time-frequency image sample set. ;in, Indicates the number of noise-adding steps Images at that time, Indicates the type of interference; from Randomly sample one In the formula, This indicates the current number of noise-adding steps. This represents the total number of noise-adding steps; Randomly generated noise In the formula, This represents Gaussian noise that follows a standard normal distribution. Calculate the noise image obtained by adding noise based on randomly generated noise and the original time-frequency image samples. ;in, In the formula, Indicates the process The image obtained by adding noise in one step, Indicates the number of noise-adding steps Images at that time, , Indicates the noise variance; The noisy image, interference type label, and noise addition step number are input into the label-guided denoising diffusion probability model to obtain the predicted noise. ; Calculate the loss function based on the randomly generated noise and the predicted noise. ; in, In the formula, Represents the loss function. Represents the L2 norm; The parameters of the label-guided denoising diffusion probability model are updated iteratively using the backpropagation algorithm. This continues until the loss function converges, resulting in a trained label-guided denoising diffusion probability model.
4. The radar active deception jamming signal identification method based on the denoising diffusion probability model according to claim 3, characterized in that, Based on the trained labels, a denoising diffusion probability model and interference type labels are used to generate time-frequency image samples, specifically including: Obtain a trained label-guided denoising diffusion probability model to determine the generated subsequences. and the interference type labels to be generated ;in, ; Randomly sample noisy images ; Based on the noise image The noise image of the previous noise-adding step is calculated. ; in, ; In the formula, Indicates the generation process The generated image obtained by adding noise in one step. It indicates the generation process. The generated image obtained by adding noise in one step. express The corresponding prediction noise, This refers to the pre-defined hyperparameters. This represents noise obtained from random sampling. express time value, express time Value; among which, , Indicates the noise variance; Iteratively calculate the noisy image from the previous noisy step until... Time-frequency image generation samples were obtained. .
5. The radar active deception jamming signal identification method based on the denoising diffusion probability model according to claim 1, characterized in that, The jamming type labels include: range spoofing jamming label, velocity spoofing jamming label, dense false target jamming label, intermittent sampling direct forwarding jamming label, intermittent sampling repeated forwarding jamming label, intermittent sampling cyclic forwarding jamming label, mixed label of range spoofing jamming and intermittent sampling repeated forwarding jamming, mixed label of range spoofing jamming and intermittent sampling direct forwarding jamming, mixed label of velocity spoofing jamming and intermittent sampling direct forwarding jamming, and mixed label of range spoofing jamming and velocity spoofing jamming.
6. The radar active deception jamming signal identification method based on the denoising diffusion probability model according to claim 1, characterized in that, The active spoofing jamming signal dataset is divided into a training set, a validation set, and a test set, specifically including: The active deception jamming signal dataset is divided into corresponding data subsets according to the jamming type label; Samples are randomly selected from each data subset according to a set ratio to form the training set, validation set, and test set.
7. The radar active deception jamming signal identification method based on the denoising diffusion probability model according to claim 6, characterized in that, The ratio of training set, validation set and test set is 7:2:
1.
8. The radar active deception jamming signal identification method based on the denoising diffusion probability model according to claim 1, characterized in that, The interference identification model can be AlexNet, support vector machine, decision tree, or logistic regression.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the radar active deception jamming signal identification method based on the denoising diffusion probability model as described in any one of claims 1-8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the radar active deception jamming signal identification method based on the denoising diffusion probability model as described in any one of claims 1-8.