Radar interference signal migration identification method, system, equipment and medium

Abstract

The invention belongs to the technical field of radars, and discloses a radar interference signal migration identification method, system and device and a medium, and the method comprises the steps: carrying out the short-time Fourier transform of an interference signal, obtaining a time-frequency diagram, and constructing an ideal interference time-frequency diagram sample set; designing a convolution attention-residual block structure, and constructing a source domain feature extractor and a target domain feature extractor; in the source domain supervised training stage, a feature extractor and a classifier are trained; in the target domain adversarial training stage, the source domain network is fixed, and a target domain feature extractor is optimized in combination with a multi-head discriminator, reconstruction loss and maximum mean difference loss; in the classification test stage, the target domain extractor and the source domain classifier are spliced to form an identification model. The scheme has better recognition performance and higher universality.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a radar interference signal migration identification method, system, equipment and medium. Background Art

[0002] With the development of modern radar technology and the advancement of electronic countermeasures, active radar jamming (AJ) has become a significant factor affecting the detection and tracking capabilities of radar systems. In complex electromagnetic environments, various active jamming signals, such as suppressive jamming, deceptive jamming, and combined jamming, can severely impact the radar system's target detection performance. To mitigate these jamming signals, radar systems must be able to accurately identify various jamming types, enabling targeted anti-jamming measures to be taken. Traditional deep learning-based AJ detection methods typically assume that the training and test set samples satisfy the independent and identically distributed (IID) condition, meaning that the feature distributions of the training and test set samples are consistent. However, in real electromagnetic environments, due to environmental factors, the feature distributions of collected non-ideal jamming signals differ significantly from those of simulated ideal jamming signals. This makes it difficult to directly apply radar jamming classification models trained under ideal conditions to non-ideal environments. Therefore, how to leverage the information from a small number of non-ideal radar jamming signal samples to improve the model's recognition capabilities for non-ideal radar jamming signals through transfer learning methods has become an urgent challenge.

[0003] Unsupervised Domain-Adversarial Adaptation (UDDA) is a transfer learning method primarily used to address domain adaptation in transfer recognition problems. This involves performing the same task in a situation where the source and target domains have different data distributions. For example, in radar jammer recognition, the training data comes from a simulation environment, while the test data comes from a real environment. This distribution discrepancy reduces the model's generalization ability. Therefore, this paper combines unsupervised domain-adversarial adaptation with deep learning to achieve transfer recognition of radar jammer signals. The following discussion focuses on the latest research progress in deep learning for unsupervised domain-adversarial adaptation.

[0004] The core idea of ​​unsupervised domain adversarial adaptation methods is to draw on the adversarial ideas of generative adversarial networks (GANs) and use adversarial training between the generator and the discriminator to narrow the feature distribution differences between source and target domain samples. Ganin et al. proposed a domain-adversarial neural network (DANN) model (GANIN Y, USTINOVA E, AJAKAN H, et al. Domain-adversarial training of neural networks[J]. Journal of machine learning research, 2016, 17(59):1-35.). This model achieves feature alignment between the source and target domains through a gradient reversal layer (GRL) and a domain discriminator, providing an efficient solution for unsupervised domain adaptation. Tzeng et al. proposed an Adversarial Discriminative Domain Adaptation (ADDA) framework. This framework minimizes the distance between the feature distributions extracted by the source domain feature extractor and the target domain feature extractor. Drawing on the two-person game idea in the GAN network, the discriminator and the target domain feature extractor are iteratively optimized to achieve feature alignment between the source and target domains (TZENG E, HOFFMAN J, SAENKO K, et al. Adversarial discriminative domain adaptation [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 7167-7176.).Volpi et al. proposed an adversarial feature augmentation method, which is an extension of the ADDA framework for adversarial discriminative domain adaptation. By training a generator to generate source-domain-like features, these features can be used to supplement source-domain training samples during adversarial training in the target domain (VOLPI R, MORERIO P, SAVARESE S, et al. Adversarial feature augmentation for unsupervised domain adaptation [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 5495-5504.). Saito et al. proposed a method based on maximum classifier discrepancy (MCD) for unsupervised domain adaptation. This method maximizes the prediction difference between the two classifiers and minimizes the difference in feature generators, making the target domain features more distinguishable, thereby improving the performance of transfer learning (SAITO K, WATANABE K, USHIKU Y, et al. Maximum classifier discrepancy for unsupervised domain adaptation [C] / / Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 3723-3732.). Peng et al. proposed an unsupervised domain adaptation method based on a diffusion model. This method gradually aligns the source and target domain data distributions through a diffusion-reverse diffusion mechanism, and combines a mutual learning strategy to promote the adaptation of the classifier to the target domain (PENG D, KE Q, AMBIKAPATHI AM, et al. Unsupervised domain adaptation via domain-adaptive diffusion [J]. IEEE Transactions on Image Processing, 2024.).Zhao et al. proposed an unsupervised domain adaptation method based on adaptive multi-scale intermediate domain and progressive training. Through multi-scale feature measurement and dynamic threshold optimization, the feature migration from the source domain to the target domain is gradually optimized. Experimental results show that this method outperforms traditional domain adaptation methods in all tasks (ZHAO X, HUANG L, NIE J, et al. Towards adaptive multi-scale intermediate domain via progressive training for unsupervised domain adaptation [J]. IEEE Transactions on Multimedia, 2023, 26: 5054-5064.).

[0005] Through the above analysis, the problems and defects of the existing technology are as follows:

[0006] (1) Existing radar interference signal recognition methods based on deep learning mainly rely on a large number of radar interference data sets obtained by software simulation under ideal conditions to train the network. However, in real non-ideal environments, the radar active interference signals received by the receiver are affected by factors such as multipath effects and noise pollution, which makes them significantly different from ideal radar interference signals. If the radar interference recognition model trained under ideal conditions is directly transferred to a non-ideal environment for interference type recognition, the recognition performance will be greatly reduced.

[0007] (2) Traditional domain adversarial adaptation methods usually perform adversarial training on the target domain feature extractor based on the classic generative adversarial network structure to achieve feature alignment between the source and target domains. However, these methods do not consider the guiding role of the source domain classification information in the training of the target domain feature extractor during the adversarial training process, which leads to low recognition performance of the radar interference classification model when it is migrated to a non-ideal environment. Summary of the Invention

[0008] In response to the problems existing in the prior art, the present invention provides a radar interference signal migration identification method, system, device and medium.

[0009] The present invention is implemented as follows: a radar interference signal migration identification method, system, device and medium comprising:

[0010] First, a short-time Fourier transform is performed on the radar interference signal to obtain a time-frequency graph of the interference signal, and a sample set of time-frequency graphs of an ideal radar interference signal is constructed. A convolutional attention-residual block structure is designed to construct the network structure of the source domain and target domain feature extractors. In the source domain supervised training phase, the source domain feature extractor and classifier are supervised and trained using the time-frequency graph samples of the source domain ideal radar interference signal, so that the source domain feature extractor can effectively extract interference signal features related to the source domain classification task. In the target domain training phase, the parameters of the source domain feature extractor and classifier are fixed, and through adversarial training between the multi-head discriminator and the target domain feature extractor, the features extracted by the target domain feature extractor are brought closer to the features extracted by the source domain feature extractor, thereby achieving feature alignment between the source domain and the target domain. In the target domain classification testing phase, the parameters of the target domain feature extractor are fixed, and the target domain feature extractor and the source domain classifier are spliced ​​to obtain a target domain non-ideal radar interference classification model, ultimately achieving the migration of the source domain (ideal radar interference) classification task to the target domain (non-ideal radar interference) classification task. The present invention has higher migration recognition accuracy and stronger versatility.

[0011] In order to achieve the above object, the technical solution adopted by the present invention is:

[0012] Step S1: Perform short-time Fourier transform on the radar jamming signal to obtain a time-frequency map, which is used as the input of the source domain and target domain feature extractor networks, and a sample set of ideal radar jamming time-frequency maps is constructed at the same time;

[0013] Step S2: Construct the source domain and target domain feature extractor network structures. In order to enhance their representation capabilities, a convolutional attention-residual block network structure is designed.

[0014] Step S3: Use the ideal radar interference time-frequency map sample set obtained in step S1 to train the source domain feature extractor and classifier network parameters.

[0015] Step S4: Fix the source domain feature extractor and classifier network parameters trained in step S3, and use the ideal radar interference time-frequency map sample set obtained in step S1 and a small number of target domain non-ideal radar interference time-frequency map sample sets to perform adversarial training on the target domain feature extractor and the improved multi-head discriminator network parameters.

[0016] Step S5: Fix the network parameters of the target domain feature extractor trained in step S4, and splice the target domain feature extractor with the source domain classifier to obtain a target domain non-ideal radar interference recognition model, so that the model can identify the radar interference type under non-ideal conditions in the target domain.

[0017] The specific method of step S1 is:

[0018] The short-time Fourier transform of the radar interference signal is expressed as:

[0019]

[0020] Where x(t) is the source signal, is the window function, STFT x (t,f) is the time-frequency diagram. Matlab software is used to simulate various radar interferences to obtain the time domain data of each interference signal, and then short-time Fourier transform is performed to obtain the source domain ideal radar interference time-frequency diagram dataset X s =(x s ,y s ∈{1,...,K}), where Represents the radar interference signal time-frequency diagram sample, y s Represents sample x s The corresponding label is the radar interference signal category corresponding to the sample, and K is the number of radar interference signal types.

[0021] The specific method of step S2 is:

[0022] The radar interference signal time-frequency diagram STFT obtained in step S1 x (t,f) is input into the source domain and target domain feature extractor network E(·) to obtain the deep abstract feature E(STFT x (t,f));

[0023] The source domain and target domain feature extraction network is constructed based on the convolutional attention mechanism-residual block structure. The convolutional attention module (CBAM) consists of a channel attention module and a spatial attention module. It adopts a step-by-step execution method to calculate the channel attention first and then the spatial attention. Then, the original input feature map is weighted to optimize the feature extraction process. Channel attention aims to learn the importance of different channels and perform weighted adjustment on each channel of the feature map. The calculation process is to extract features from the time-frequency map of the radar interference signal through the residual network to obtain a feature map matrix containing multiple channels. Then use global average pooling and global maximum pooling to extract the global information of the feature map in the channel dimension to obtain F avg and F max , which is calculated as follows:

[0024]

[0025] Then the results of global average pooling and global maximum pooling are input into the shared MLP (multi-layer perceptron) to obtain the channel attention weight M c(F), which is calculated as follows:

[0026] M c (F)=σ(W2(δ(W1(F avg )))+W2(δ(W1(F max ))))

[0027] Among them, MLP consists of two layers of fully connected layers, W1 and W2 are the parameters of the first and second fully connected layers respectively, which are used to learn the dependencies between channels, δ represents the ReLU activation function, and σ represents the Sigmoid activation function, which is used to map the weights to [0,1]. Finally, the attention weight M c (f) Act on the input feature map F to obtain the weighted feature map F′ and complete the channel weighting. The calculation formula is as follows:

[0028] F′=M c (F)×F

[0029] Different from channel attention, spatial attention aims to learn the importance of different spatial positions in the feature map, so that the network can focus on the areas that play a key role in the recognition task. Its calculation process is to perform maximum pooling and average pooling on the input feature map in the channel dimension to obtain spatial information. The calculation formula is as follows:

[0030]

[0031] in, and Represent the feature maps after channel average pooling and maximum pooling, respectively, with a size of H×W. Then, these two feature maps are spliced ​​in the channel dimension and further dimensionally compressed using a 7×7 convolution layer to obtain the spatial attention weight M s (F), which is calculated as follows:

[0032]

[0033] Among them, σ is the Sigmoid activation function, Conv 7×7 The 7×7 convolution layer is used to learn the importance of spatial dimensions. Finally, the calculated spatial attention weight M s (F) acts on the feature map F′ after the channel attention module weighting to obtain the final weighted output feature map F″ to complete the spatial weighting. The calculation formula is as follows:

[0034] F″=M s (F′)×F′

[0035] The specific method of step S3 is:

[0036] In the source domain supervised training phase, we first use the labeled source domain sample set (ideal radar interference signal time-frequency graph dataset) X in a supervised learning manner. S =(x s ,y s ∈{1,...,K}) for the source domain feature extractor E S and classifier C s For training, the optimization goal is to minimize the cross entropy loss function of K categories as shown below:

[0037]

[0038] After training is completed, the source domain feature extractor E is fixed s and classifier C s The parameters of , thus fixing the feature distribution H after the source domain sample set is input into the source domain feature extractor s , then the source domain feature distribution H s As a reference, the target feature extractor E is obtained through adversarial training. T , make the target domain feature distribution H as much as possible T and the source domain feature distribution H s Stay consistent.

[0039] The specific method of step S4 is:

[0040] In the target domain adversarial training process, the discriminator loss function mainly consists of two parts: source domain reconstruction loss and target domain discrimination loss The source domain reconstruction loss L D,s Posterior probability of source domain feature extractor set to zero splicing and the discriminator posterior probability q s =D(h s ), the cross entropy between them is calculated as follows:

[0041]

[0042] Among them, p s,k and q s,k They are and q s It is worth noting that the kth element of is represented by splicing a zero into the posterior probability of the source domain feature extractor to represent the K+1th target domain category, thereby keeping the K+1-dimensional probability distribution output by the discriminator summed to 1.

[0043] Target domain discrimination loss The optimization goal is to output h for the target domain feature extractor t, the discriminator should judge it as a target domain sample, that is, output the K+1th class, and its calculation formula is:

[0044]

[0045] Among them, q t (K+1) is the probability of the target domain category output by the discriminator.

[0046] The loss function of the target domain feature extractor also consists of two parts: the discriminant loss of the target domain and the loss of the target domain. and maximum mean difference loss Among them, the target domain discrimination loss The optimization goal is to minimize the probability that the discriminator classifies the output of the target domain feature extractor as the K+1th class, even if the discriminator cannot effectively distinguish whether the sample comes from the source domain or the target domain. The calculation formula is as follows:

[0047]

[0048] Maximum mean difference loss The optimization goal is to hope that the output generated by the target domain feature extractor can be identified by the discriminator as one of the first K specific task categories, rather than the K+1 target domain category, so that the output of the target domain feature extractor and the output of the source domain feature extractor are as close as possible in the feature space. The specific implementation method is to convert the source domain sample x s Input to the source domain feature extractor to obtain the source domain feature h s , and then h s Input into the classifier and discriminator respectively to generate two source domain posterior probabilities p s =C s (h s ) and q s =D(h s ). Unlike supervised domain adaptation methods, since the target domain sample set does not contain labels, the source domain and target domain samples cannot form a paired correspondence, and it is impossible to perform paired tests on these posterior probabilities. Therefore, it is assumed here that the distribution of the target domain posterior probability q t =D(h t ) and the distribution of the source domain posterior probability p s are equal, and the two are compared, and a set of target domain samples and source domain samples are input into the network to obtain a set of target domain posterior probabilities and a set of source domain posterior probabilities At this point, the optimization objective is transformed into minimization and The distance between them.

[0049] The present invention uses the Maximum Mean Discrepancy (MMD) to measure the source domain posterior distribution p s and the target domain posterior distribution q s The mean distance between MMD measures the difference between distributions by mapping the data into the reproducing kernel Hilbert space (RKHS). The specific calculation formula is as follows:

[0050]

[0051] Where k(·,·) is a kernel function. Usually, Gaussian kernel is selected as the metric function, which is defined as:

[0052]

[0053] Where σ is the standard deviation of the Gaussian kernel, which determines the width of the kernel function. Ultimately, the maximum mean difference loss It can be defined as:

[0054]

[0055] The specific method of step S5 is:

[0056] After the target domain feature extractor is trained, the target domain feature extractor and the source domain classifier are spliced ​​and directly used for target domain data (non-ideal radar interference) x t Classification and recognition, when reasoning, use the target domain feature extractor E S Extract features and finally input them into the classifier C S The target domain radar interference signal type can be output in , and its calculation process can be expressed as:

[0057]

[0058] A system for a radar interference signal migration identification method, characterized by comprising:

[0059] A short-time Fourier transform module is used to perform short-time Fourier transform on the radar interference signal in step S1, and the obtained time-frequency map is used as the input of the source domain and target domain feature extractor network, and to construct an ideal radar interference time-frequency map sample set;

[0060] The source domain and target domain feature extractor based on the convolutional attention-residual block structure is used to perform deep feature extraction on the radar interference signal time-frequency diagram obtained in step S1 using the source domain and target domain feature extractor based on the convolutional attention-residual block structure in step S2.

[0061] The adversarial training stage of the target domain feature extractor based on the multi-head discriminator is used to perform adversarial training on the target domain feature extractor network parameters constructed in step S2 in step S4, and design the source domain reconstruction loss function and the maximum mean difference loss function.

[0062] A device for a radar interference signal migration identification method includes: a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, any one of the above radar interference signal migration identification methods can be implemented.

[0063] A computer storage medium for receiving a program input by a user, wherein the computer program stored in the storage medium, when executed by a processor, can perform radar interference signal migration identification based on the radar interference signal migration identification method described in any one of the above technical solutions.

[0064] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0065] First, the present invention proposes a source domain and target domain feature extractor network structure based on the convolutional attention mechanism-residual block structure. Combining the convolutional attention module (CBAM) with the residual block structure in ResNet can enable the feature extractor to focus more on capturing the areas in the radar interference time-frequency graph that play a key role in the classification task, thereby improving the feature discrimination of the feature extractor in extracting radar interference signals and further improving the transfer recognition performance of the model.

[0066] The present invention expands the output dimension of the discriminator in the traditional adversarial discriminant domain adaptation method, expanding its output dimension from 2 dimensions to dimensions. At the same time, a source domain reconstruction loss function and a maximum mean difference loss function are designed to perform adversarial training on the target domain feature extractor. In this way, the discriminator no longer focuses solely on domain distinction, but can also model different categories in the source domain, thereby better guiding the target domain feature extractor to learn feature representations related to the classification task during the adversarial training process, further improving the model's ability to recognize different types of radar active jammers.

[0067] Compared with the existing technology, the present invention studies and explores the importance of constructing source domain and target domain feature extractor networks based on the convolutional attention-residual block structure and adversarial training of target domain feature extractors based on a multi-head discriminator in radar interference signal migration identification; any scenario involving radar interference signal migration identification, especially when there is a significant difference in the feature distribution of the non-ideal radar interference sample set collected in a real environment and the ideal radar interference sample set generated by the software simulation used for model training, can be used for migration identification; the source domain and target domain feature extractors based on the convolutional attention-residual block structure can enable the network to effectively focus on the key areas in the radar interference time-frequency diagram through their spatial attention and channel attention, and the residual structure is used to alleviate the gradient disappearance phenomenon; the adversarial training of the target domain feature extractor based on the multi-head discriminator expands the output dimension of the traditional discriminator and introduces source domain classification task information to assist training, thereby solving the problem of classification performance degradation caused by the loss of source domain classification task information in the target domain feature extractor training process, and further improving the recognition performance after model migration; the present invention has higher recognition accuracy and stronger versatility.

[0068] This paper addresses the limited feature extraction capabilities and inter-domain distribution shift that lead to reduced recognition accuracy in existing radar jamming signal recognition technologies. By proposing a feature extraction network based on a convolutional attention-residual block structure, this approach combines a multi-headed discriminator adversarial training mechanism with a joint loss function optimization strategy. After obtaining the time-frequency map of the radar jamming signal using a short-time Fourier transform, the attention mechanism is used to enhance modeling capabilities for regions with significant jamming features. Furthermore, a source domain reconstruction loss and a maximum mean difference loss are introduced to guide target domain features to align with the source domain distribution, improving the model's adaptability and robustness under non-ideal target domain conditions.

[0069] Compared with existing single-discriminator architectures and shallow feature extraction methods, this method achieves significant technical advancements in algorithm design and mathematical model construction. The multi-head discriminant architecture provides multi-angle discrimination capabilities, avoiding overfitting of source domain feature distributions. Maximum mean difference optimization effectively reduces cross-domain distribution differences, significantly improving the accuracy and stability of transfer learning. Experimental results demonstrate that this method outperforms existing mainstream methods in multiple radar jamming signal recognition tasks, demonstrating enhanced generalization and recognition performance in heterogeneous environments.

[0070] Second, the radar interference transfer recognition method based on a multi-head discriminator proposed in this paper expands the discriminator output dimension in the traditional ADDA method and introduces information from the source domain classification task to guide the target domain feature extractor for adversarial training. This solves the problem in traditional methods where the target domain feature extractor loses its ability to classify and recognize source domain radar interference categories during training due to the lack of source domain classification information. Furthermore, by designing a feature extractor structure based on a convolutional attention-residual block structure, the radar interference classification and recognition performance under real-world non-ideal conditions is further improved. In the field of radar interference type recognition, previous deep learning-based radar interference recognition methods typically supervise the training of classification models based on simulated radar interference signal data. However, in real-world electromagnetic environments, radar interference signals are affected by environmental factors, making the classification models trained by traditional methods incapable of direct application in real-world environments, resulting in low practical application value. The proposed method improves the interference recognition performance of radar interference classification models in real-world non-ideal electromagnetic environments, making it widely applicable to radar interference recognition in real battlefield electromagnetic environments. It can effectively improve the anti-interference capability of our military radar systems, thereby ensuring that our radar systems can properly detect and locate targets. If the present invention is industrialized, it can be widely used in military radar anti-interference systems. At the same time, the method has good deployability and can be quickly deployed on machines. It can bring considerable economic benefits to radar system manufacturers and effectively improve the anti-interference performance of radar systems in real non-ideal electromagnetic environments, thereby enhancing the market competitiveness of products.

[0071] At present, the research in the field of radar interference migration recognition at home and abroad mainly focuses on the direction of fine-tuning migration, that is, using transfer learning technology to fine-tune the pre-trained classification model with a small number of radar interference samples, so as to realize the recognition of radar interference types under small sample conditions. However, although this migration method solves the problem of low model recognition performance caused by the insufficient number of radar interference samples available for training in real situations, the radar interference signals in the real electromagnetic environment are affected by environmental factors, resulting in large differences in features between them and the radar interference samples in the training set, which makes the trained radar interference classification model unable to be directly applied to the real electromagnetic environment. The present invention effectively realizes the feature alignment between the ideal radar interference signal in the source domain training set and the actual non-ideal radar interference signal in the target domain through adversarial training between the target domain feature extractor and the discriminator, thereby realizing the migration of the model in the real non-ideal environment, improving the migration recognition performance of the model, and filling the gap in radar interference migration recognition technology in non-ideal environments at home and abroad. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is a flow chart of the radar interference signal migration identification method, system, medium and equipment provided by an embodiment of the present invention.

[0073] Figure 2 This is a convolutional attention-residual block network structure diagram in the radar interference signal migration identification method provided by an embodiment of the present invention.

[0074] Figure 3 It is a schematic diagram of the workflow of the source domain supervised training process in the radar interference signal migration identification method provided by an embodiment of the present invention.

[0075] Figure 4 This is a schematic diagram of the workflow of the adversarial training process of the target domain feature extractor based on the multi-head discriminator in the radar interference signal migration identification method provided by an embodiment of the present invention.

[0076] Figure 5 This is a graph showing experimental results comparing the recognition rates of the radar interference signal migration identification method provided by an embodiment of the present invention before and after migration for different radar interference categories.

[0077] Figure 6 This is a graph showing the comparison of migration recognition rate experimental results of the radar interference signal migration recognition method provided by an embodiment of the present invention under different source domain and target domain feature extractor network structures.

[0078] Figure 7 This is a graph showing the experimental results of comparing the recognition performance of the radar interference signal migration recognition method provided by an embodiment of the present invention with other migration recognition methods. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0080] In response to the problems existing in the prior art, the present invention provides a radar interference signal migration identification method, medium and device. The present invention is described in detail below with reference to the accompanying drawings.

[0081] Radar interference signal migration identification method, system, equipment and medium, the method includes: first, performing short-time Fourier transform on the radar interference signal to obtain the interference signal time-frequency diagram, and constructing an ideal radar interference signal time-frequency diagram sample set; designing a convolutional attention-residual block structure to construct the source domain and target domain feature extractor network structure; in the source domain supervision training stage, using the source domain ideal radar interference time-frequency diagram sample set to train the source domain feature extractor and source domain classifier; in the target domain adversarial training stage, designing the source domain reconstruction loss and maximum mean difference loss function, fixing the source domain feature extractor and source domain classifier parameters, using the target domain non-ideal radar interference signal time-frequency diagram sample set and the source domain ideal radar interference signal time-frequency diagram sample set with the help of a multi-head discriminator to perform adversarial training on the target domain feature extractor; in the target domain classification test stage, fixing the target domain feature extractor parameters, and splicing the target domain feature extractor and the source domain classifier to obtain a target domain non-ideal radar interference identification model. The present invention has better performance and wider versatility; the system, equipment and medium are used to implement the radar interference signal migration identification method

[0082] The radar interference signal migration identification method, system, medium and equipment provided by the present invention can also be implemented by ordinary technicians in other steps. Figure 1 The radar interference signal migration identification method, medium and device provided by the present invention are only a specific embodiment.

[0083] like Figure 1 As shown, the radar interference signal open set identification method provided by the embodiment of the present invention has the following specific steps:

[0084] Step S1: Perform short-time Fourier transform on the radar interference signal to obtain a time-frequency map. The obtained time-frequency map is used as the input of the source domain and target domain feature extractor network, and an ideal radar interference time-frequency map sample set is constructed at the same time. The specific process is as follows:

[0085] The short-time Fourier transform of the radar interference signal is expressed as:

[0086]

[0087] Where x(t) is the source signal, is the window function, STFT x (t,f) is the time-frequency diagram. Matlab software is used to simulate various radar interferences to obtain the time domain data of each interference signal, and then short-time Fourier transform is performed to obtain the source domain ideal radar interference time-frequency diagram dataset X s =(x s ,y s ∈{1,...,K}), where Represents the radar interference signal time-frequency diagram sample, y sRepresents sample x s The corresponding label is the radar interference signal category corresponding to the sample, and K is the number of radar interference signal types.

[0088] Step S2: Construct the source domain and target domain feature extractor network structures. In order to enhance their representation capabilities, a convolutional attention-residual block network structure is designed. The specific process is as follows:

[0089] The radar interference signal time-frequency diagram STFT obtained in step S1 x (t,f) is input into the source domain and target domain feature extractor network E(·) to obtain the deep abstract feature E(STFT x (t,f));

[0090] The source domain and target domain feature extraction network is constructed based on the convolutional attention mechanism-residual block structure. The convolutional attention module (CBAM) consists of a channel attention module and a spatial attention module. It adopts a step-by-step execution method to calculate the channel attention first and then the spatial attention. Then, the original input feature map is weighted to optimize the feature extraction process. Channel attention aims to learn the importance of different channels and perform weighted adjustment on each channel of the feature map. The calculation process is to extract features from the time-frequency map of the radar interference signal through the residual network to obtain a feature map matrix containing multiple channels. Then use global average pooling and global maximum pooling to extract the global information of the feature map in the channel dimension to obtain F avg and F max , which is calculated as follows:

[0091]

[0092] Then the results of global average pooling and global maximum pooling are input into the shared MLP (multi-layer perceptron) to obtain the channel attention weight M c (F), which is calculated as follows:

[0093] M c (F)=σ(W2(δ(W1(F avg )))+W2(δ(W1(F max ))))

[0094] Among them, MLP consists of two layers of fully connected layers, W1 and W2 are the parameters of the first and second fully connected layers respectively, which are used to learn the dependencies between channels, δ represents the ReLU activation function, and σ represents the Sigmoid activation function, which is used to map the weights to [0,1]. Finally, the attention weight M c(f) Act on the input feature map F to obtain the weighted feature map F′ and complete the channel weighting. The calculation formula is as follows:

[0095] F′=M c (F)×F

[0096] Different from channel attention, spatial attention aims to learn the importance of different spatial positions in the feature map, so that the network can focus on the areas that play a key role in the recognition task. Its calculation process is to perform maximum pooling and average pooling on the input feature map in the channel dimension to obtain spatial information. The calculation formula is as follows:

[0097]

[0098] in, and Represent the feature maps after channel average pooling and maximum pooling, respectively, with a size of H×W. Then, these two feature maps are spliced ​​in the channel dimension and further dimensionally compressed using a 7×7 convolution layer to obtain the spatial attention weight M s (F), which is calculated as follows:

[0099]

[0100] Among them, σ is the Sigmoid activation function, Conv 7×7 The 7×7 convolution layer is used to learn the importance of spatial dimensions. Finally, the calculated spatial attention weight M s (F) acts on the feature map F′ after the channel attention module weighting to obtain the final weighted output feature map F″ to complete the spatial weighting. The calculation formula is as follows:

[0101] F″=M s (F′)×F′

[0102] Convolutional attention mechanism - residual block structure such as Figure 2 As shown in the figure, channel attention can dynamically adjust the weight of each feature map channel to highlight the characteristic pattern of radar interference signals; spatial attention can enable the network to pay more attention to areas with abnormal or sudden energy distribution on the time-frequency graph, thereby more effectively distinguishing different types of radar interference signals.

[0103] Step S3: Use the ideal radar interference time-frequency map sample set obtained in step S1 to train the source domain feature extractor and classifier network parameters. The specific process is as follows:

[0104] like Figure 3 As shown in the source domain supervision training stage, firstly, the source domain sample set with labels (ideal radar interference signal time-frequency diagram dataset) X is used in a supervised learning manner.S =(x s ,y s ∈{1,...,K}) for the source domain feature extractor E S and classifier C s For training, the optimization goal is to minimize the cross entropy loss function of K categories as shown below:

[0105]

[0106] After training is completed, the source domain feature extractor E is fixed s and classifier C s The parameters of , thus fixing the feature distribution H after the source domain sample set is input into the source domain feature extractor s , then the source domain feature distribution H s As a reference, the target feature extractor E is obtained through adversarial training. T , make the target domain feature distribution H as much as possible T and the source domain feature distribution H s Stay consistent.

[0107] Step S4: Fix the source domain feature extractor and classifier network parameters trained in step S3, and use the ideal radar interference time-frequency map sample set obtained in step S1 and a small number of target domain non-ideal radar interference time-frequency map sample sets to perform adversarial training on the target domain feature extractor and the improved multi-head discriminator network parameters. The specific process is as follows:

[0108] like Figure 4 As shown in the figure, during the target domain adversarial training process, the discriminator's loss function mainly consists of two parts: the source domain reconstruction loss and and target domain discrimination loss The source domain reconstruction loss L D,s Posterior probability of source domain feature extractor set to zero splicing and the discriminator posterior probability q s =D(h s ), the cross entropy between them is calculated as follows:

[0109]

[0110] Among them, p s,k and q s,k They are and q s It is worth noting that the kth element of is represented by splicing a zero into the posterior probability of the source domain feature extractor to represent the K+1th target domain category, thereby keeping the K+1-dimensional probability distribution output by the discriminator summed to 1.

[0111] Target domain discrimination loss The optimization goal is to output h for the target domain feature extractor t , the discriminator should judge it as a target domain sample, that is, output the K+1th class, and its calculation formula is:

[0112]

[0113] Among them, q t (K+1) is the probability of the target domain category output by the discriminator.

[0114] The loss function of the target domain feature extractor also consists of two parts: the discriminant loss of the target domain and the loss of the target domain. and maximum mean difference loss Among them, the target domain discrimination loss The optimization goal is to minimize the probability that the discriminator classifies the output of the target domain feature extractor as the K+1th class, even if the discriminator cannot effectively distinguish whether the sample comes from the source domain or the target domain. The calculation formula is as follows:

[0115]

[0116] Maximum mean difference loss The optimization goal is to hope that the output generated by the target domain feature extractor can be identified by the discriminator as one of the first K specific task categories, rather than the K+1 target domain category, so that the output of the target domain feature extractor and the output of the source domain feature extractor are as close as possible in the feature space. The specific implementation method is to convert the source domain sample x s Input to the source domain feature extractor to obtain the source domain feature h s , and then h s Input into the classifier and discriminator respectively to generate two source domain posterior probabilities p s =C s (h s ) and q s =D(h s ). Unlike supervised domain adaptation methods, since the target domain sample set does not contain labels, the source domain and target domain samples cannot form a paired correspondence, and it is impossible to perform paired tests on these posterior probabilities. Therefore, it is assumed here that the distribution of the target domain posterior probability q t =D(h t ) and the distribution of the source domain posterior probability p s are equal, and the two are compared, and a set of target domain samples and source domain samples are input into the network to obtain a set of target domain posterior probabilities and a set of source domain posterior probabilities At this point, the optimization objective is transformed into minimization and The distance between them.

[0117] The present invention uses the Maximum Mean Discrepancy (MMD) to measure the source domain posterior distribution p s and the target domain posterior distribution q s The mean distance between MMD measures the difference between distributions by mapping the data into the reproducing kernel Hilbert space (RKHS). The specific calculation formula is as follows:

[0118]

[0119] Where k(·,·) is a kernel function. Usually, Gaussian kernel is selected as the metric function, which is defined as:

[0120]

[0121] Where σ is the standard deviation of the Gaussian kernel, which determines the width of the kernel function. Ultimately, the maximum mean difference loss It can be defined as:

[0122]

[0123] Step S5: Fix the network parameters of the target domain feature extractor trained in step S4, and combine the target domain feature extractor with the source domain classifier to obtain a target domain non-ideal radar interference recognition model, so that the model can identify the radar interference type under non-ideal conditions in the target domain. The specific process is as follows:

[0124] After the target domain feature extractor is trained, the target domain feature extractor and the source domain classifier are spliced ​​and directly used for target domain data (non-ideal radar interference) x t Classification and recognition, when reasoning, use the target domain feature extractor E S Extract features and finally input them into the classifier C S The target domain radar interference signal type can be output in , and its calculation process can be expressed as:

[0125]

[0126] In order to evaluate the performance of the present invention, simulation verification was carried out. The specific simulation conditions and parameters are as follows:

[0127] The experiments were conducted using a computer with a 16-core AMD EPYC 7502 processor, an NVIDIA RTX 4090 GPU, and 96GB of RAM. The operating system was Ubuntu 20.04, the software environment was Python 3.10, and the deep learning framework Pytorch 2.3.0 was used for network construction and training. During training, the learning rate was set to 0.002, the batch size was 64, the number of epochs was 100, and a stochastic gradient descent optimizer with a momentum factor of 0.9 was used. The experiments focused on eight types of radar active jammers: noise amplitude modulation (AM), noise frequency modulation (FM), comb spectrum jammer (COMB), spectrum dispersion jammer (SMSP), slice reconstruction jammer (CI), intermittent sampling and forwarding jammer (IS), densely replicated false target jammer (MT), and suppression-deception combined jammer. The sampling frequency is set to 2.4 GHz. Taking the source domain sample set as an example, the interference-to-noise ratio range is set to 0–8 dB with an interval of 2 dB. For each interference-to-noise ratio in the training set, 300 time-frequency graph samples are generated, resulting in a total of 8 × 5 × 300 = 12,000 samples. For each interference-to-noise ratio in the test set, 60 time-frequency graph samples are generated, resulting in a total of 8 × 5 × 60 = 2,400 samples.

[0128] The source and target domain sample sets were divided into training and test sets. Following the CIFAR-10 image dataset's partitioning criteria, the training and test sets were generated in a 5:1 ratio. To investigate the impact of the number of target domain training samples on radar active jammer migration recognition performance, six target domain training sets with varying sample sizes were created: 600, 1200, 2400, 3600, 4800, and 6000. The target domain non-ideal conditions were divided into two types: target domain 1 and target domain 2. The non-ideal conditions in target domain 1 were point-frequency noise, while those in target domain 2 were strongly suppressive noise (simulated using Gaussian white noise). The number of point-frequency signals was randomly selected between [20, 40], and the frequency range of the point-frequency signals was [0.75f0, 1.25f0]. f0 was the center frequency of the radar jammer, and the interference-to-noise ratio (INR) between the radar jammer and the strongly suppressive noise was set to -5dB.

[0129] Figure 5 The following figure shows the comparison of model recognition performance before and after migration under different radar interference categories. Figure 5It can be seen that compared with before migration, the recognition rates of most interference types have been improved after migration, and the recognition rates of all radar interference types are higher than 80%, among which the recognition rates of AM, FM, COMB, SMSP, IS and MT are higher than 90%. On the other hand, since the method proposed in the present invention enables the model to better learn the characteristic distribution of non-ideal radar interference signals in the target domain through domain adversarial training, thereby reducing domain offset, the recognition effect of each interference category in the target domain is more stable after migration. Experimental results show that the method proposed in the present invention can effectively achieve feature alignment between the source domain and the target domain through adversarial training between the target domain feature extractor and the multi-head discriminator, thereby improving the model's recognition performance for radar interference under non-ideal conditions. At the same time, the method proposed in the present invention can maintain a high recognition accuracy for the eight types of radar active interference studied.

[0130] Figure 6 The following figure shows the comparison of model recognition performance under different source domain and target domain feature extractor network structures. In order to verify the effectiveness of the convolutional attention mechanism-residual block (CBAM-ResConv) network structure adopted by the source domain feature extractor and the target domain feature extractor in the method proposed in this paper, a control experiment was set up to replace the CBAM-ResConv layer in the feature extractor of this paper with the residual convolution structure (ResConv) without the convolutional attention mechanism and the simple convolution layer (Conv) structure, while the other network layer structures remain unchanged. The migration recognition performance of the model under different network structures is tested, and the experimental results are shown in Figure 2. Figure 6 As shown in the experimental results, the transfer recognition rate of the three feature extractor network structures increases with the number of target domain samples, and the recognition performance of the CBAM-ResConv network structure adopted in this paper is consistently better than the other two network structures. The experimental results show that the convolutional attention mechanism residual module CBAM-ResConv can effectively capture the key information in the time-frequency map features of the radar jamming signal input through its channel attention and spatial attention mechanisms, thereby improving the model transfer recognition performance.

[0131] Figure 7The comparison chart of the migration recognition performance of the method proposed in the present invention and other migration recognition methods is shown. In order to verify that the method in this paper has better recognition performance for the radar interference signal migration recognition task under non-ideal conditions than other domain adversarial adaptive methods, three methods, DANN, ADDA and CDAN, are selected for comparative experiments. Among them, the DANN method uses a shared feature extractor in the source domain and the target domain, and performs adversarial training through the gradient reversal layer (GRL), so as to achieve feature alignment between the source domain and the target domain. The ADDA method is similar to the method in this paper. It also adopts a two-stage training method and uses adversarial training between the target domain feature extractor and the source domain feature extraction to achieve feature alignment. The CDAN method is an improvement on DANN. It introduces category information on the basis of DANN to enhance the alignment ability of domain features. In order to ensure the fairness of the comparative experiment, the source domain and target domain feature extractors, classifiers, and discriminators of the three comparative methods adopt the same network structure as the method proposed in this invention. The migration recognition performance of different methods was tested under two non-ideal conditions, target domain 1 and target domain 2, respectively. The experimental results are shown in the figure below. Figure 7 As shown in the figure, the recognition rates of the four methods all increase with the increase in the number of target domain samples, and are significantly improved compared to the pre-transfer results. This shows that the four methods can achieve effective transfer for both non-ideal cases of target domain 1 and target domain 2.

[0132] For target domain 1 and target domain 2, the transfer recognition effect of the proposed algorithm is better than that of the other three algorithms under different numbers of target domain samples. This is because the source domain and target domain feature extractors are shared in the DANN and CDAN methods. When the number of target domain samples is small, due to insufficient adversarial training, the feature extractor is more adapted to the feature distribution of the source domain. In the proposed method, the target domain feature extractor is independent and not dominated by the source domain. Therefore, the transfer recognition performance of these two methods is lower than that of the proposed method. Although the ADDA method has an independent target domain feature extractor like the proposed method, it does not introduce the classification information of the source domain task during the adversarial training process. As the adversarial process deepens, the model may pay more attention to the differences between the source domain and target domain features and ignore the classification task itself, resulting in its transfer recognition performance being lower than that of the proposed method.

[0133] It is worth noting that compared with target domain 1, the method proposed in this paper has relatively obvious advantages over other methods in target domain 2. This is because the method proposed in this paper expands the output dimension of the discriminator and introduces the source domain classification task information into the training by designing the maximum mean difference loss. In this way, when the target domain feature extractor faces non-ideal situations with large feature differences such as target domain 2, it can use the source domain classification information to more accurately align features to specific categories in the source domain.

[0134] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0135] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A radar interference signal migration identification method, characterized in that: The following steps are involved: S1. Perform short-time Fourier transform on the radar interference signal to obtain a time-frequency diagram and construct an ideal radar interference time-frequency diagram sample set; S2. Based on the convolutional attention-residual block structure, the source domain feature extractor and target domain feature extractor networks are constructed respectively; S3. Use the ideal radar interference time-frequency map sample set to supervise the source domain feature extractor and classifier to obtain fixed network parameters; S4. Fix the source domain network parameters, combine the target domain non-ideal samples with the source domain samples, and perform adversarial training on the target domain feature extractor through the source domain reconstruction loss and the maximum mean difference loss; S5. Combine the trained target domain feature extractor with the source domain classifier to build a target domain recognition model to realize the recognition of radar interference types under non-ideal conditions.

2. The radar interference signal migration identification method according to claim 1, wherein: In step S2, the convolutional attention-residual block structure includes a channel attention module and a spatial attention module. The channel attention module obtains channel weights through global average pooling and maximum pooling. The spatial attention module obtains the weights of the spatial dimension through convolution operations and performs two-level weighted optimization on the feature map.

3. The radar interference signal migration identification method according to claim 1, wherein: In step S3, the source domain classifier is trained and optimized using a cross entropy loss function to minimize the difference between the predicted label and the true label.

4. The radar interference signal migration identification method according to claim 1, wherein: In step S4, the source domain reconstruction loss is the cross entropy between the source domain posterior probability and the discriminator output probability; the maximum mean difference loss is used to measure the mean distance between the source domain and the target domain posterior probabilities, and a Gaussian kernel is used for mapping and calculation.

5. The radar interference signal migration identification method according to claim 1, wherein: A multi-head discriminator structure is adopted in the target domain adversarial training, and each discriminator is used to identify specific radar interference features, thereby improving the generalization ability of target domain recognition.

6. A radar interference signal migration identification system implementing the radar interference signal migration identification method according to any one of claims 1 to 5, characterized in that: The radar interference signal migration identification system includes: Short-time Fourier transform module, used to transform the input radar interference signal to obtain a time-frequency diagram; Feature extraction module, including source domain and target domain feature extractors, built on the convolutional attention-residual structure; The training module is used to perform source domain supervised training and target domain adversarial training, including the discriminator and classifier; A concatenated inference module is used to combine the target domain feature extractor with the source domain classifier for final recognition.

7. A radar interference signal identification device, characterized in that: include: A processor and a memory, wherein the memory stores a computer program, and when the program is executed by the processor, the radar interference signal migration identification method according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium having program instructions stored thereon, wherein when the program instructions are executed by a processor, the method according to any one of claims 1 to 5 is executed.

9. A feature extraction structure for radar interference signal identification implementing the radar interference signal migration identification method according to any one of claims 1 to 5, characterized in that: The structure includes a convolutional neural network, a channel attention module and a spatial attention module. The channel attention module weights the channel features through a multi-layer perceptron, and the spatial attention module weights the spatial dimension through pooling and convolution to achieve the optimal abstract expression of the interference signal feature map.

10. A radar interference migration discrimination structure for domain adaptation implementing the radar interference signal migration identification method according to any one of claims 1 to 5, characterized in that: The method includes a fixed source domain feature extraction network, a target domain feature extraction network and multiple discriminators; the discriminator receives the output from the target domain network and jointly trains the target domain feature extraction network based on the target domain discrimination loss and the maximum mean difference loss to improve its adaptability to target domain interference signals.

Images