Non-identical distribution radar jamming recognition method based on adversarial transfer learning

Abstract

The present invention provides a non-identical distribution radar interference identification method based on adversarial transfer learning, comprising the following steps: first, generating an ideal interference signal; then generating a source domain sample set consisting of ideal interference signal time-frequency graph samples, and a target domain sample set consisting of non-identical distribution interference signal time-frequency graph samples simulated and added with non-ideal factors; then, through the radar interference identification step based on adversarial transfer learning, which includes three processes: source domain pre-learning, adversarial discrimination, and target domain classification test, obtaining a target domain convolutional network that has completed adversarial discrimination training and a classifier that has completed supervised interference signal identification training, inputting a signal to be identified into the target domain convolutional network for feature extraction and then outputting it to the classifier, which completes the final radar interference signal identification and classification and outputs the final classification recognition result. The present invention can achieve both the computational efficiency and generalization of the network, has a high recognition rate, and can effectively classify non-identical distribution samples.

Description

Technical Field

[0001] The present invention relates to radar interference recognition technology, in particular to non-uniformly distributed radar interference recognition technology. Background Art

[0002] In recent years, with the continuous development of digital radio frequency memory jammer (DRFM) technology, jammers can complete signal interception, modulation and forwarding of jamming signals in a very short time. The active jamming signals generated by this technology have strong coherence characteristics, which makes it difficult to distinguish them from radar echo signals, posing a serious threat to the effective operation of radars.

[0003] While radar jamming technology continues to rapidly advance, research on signal recognition, classification, and countermeasures in complex active jamming environments has lagged behind. Jamming identification research lacks a unified and effective method for identifying common active jamming signals. Jamming identification and parameter estimation rely heavily on operator judgment and the hardware settings of the receiving device, resulting in high levels of subjectivity and ambiguity.

[0004] A traditional radar interference recognition method is an interference recognition algorithm based on the maximum likelihood criterion. This algorithm uses a probability model to determine the type of interference. See: Li Jinzong, Wei Xiangquan, Li Ningning, and Wang Song. Point Target Recognition Technology Based on the Maximum Likelihood Ratio Criterion [J]. Journal of Electronics, 2003(08):1217-1221. This interference recognition algorithm based on the maximum likelihood criterion is based on prior information obtained from subjective experience and relies on the observation of the radar operator in its judgment. It has poor engineering applicability and the detection results are not reliable.

[0005] The support vector machine (SVM) decision classification method for radar jamming signal classification uses decision tree theory to construct a series of two-class classifiers, each of which classifies a certain input sample into one category and the remaining input samples into another category. See: S. Hassanpour, AM Pezeshk, F. Behnia. Automatic digital modulation recognition based on novel features and support vector machine [C]. IEEE International Conference on Signal-Image Technology & Internet-Based Systems, Naples, Italy, 2017, 172-177. The SVM method solves the problem of determining the number of hidden layer nodes in neural network classification methods. However, choosing an appropriate kernel function that ensures both computational efficiency and good generalization is a challenge. It is not possible to achieve both computational efficiency and generalization at the same time.

[0006] Deep learning methods are used to extract the features of the transformed ideal interference samples and automatically classify them based on the feature space. See: Liu Qiang. Radar Interference Identification Technology Based on Deep Learning [D]. University of Electronic Science and Technology of China, 2020. Using convolutional neural networks for interference signal identification improves signal feature extraction and classification efficiency, but its generalization is weak and it cannot effectively identify non-identically distributed interference. Ideal interference here refers to interference signals in an ideal scenario without background noise, clutter, or other superimposed interference signals.

[0007] In complex communication environments, various non-ideal factors exist in the generation, transmission, and reception processes of interference, affecting the characteristics of various domains of the interference signal. Compared to ideal scenarios, radar interference signals in real-world scenarios exhibit significantly different parameters and characteristics, and are more likely to encounter complex situations involving the superposition of multiple interference signals. The distribution of interference in ideal and real-world scenarios differs. Interference signals in real-world scenarios exhibit non-identical distributions.

[0008] The existing traditional signal recognition and classification methods based on likelihood ratio, support vector machine, and convolutional neural network are not general enough, and their classification models are unable to effectively identify and classify interference signal samples with large distribution differences. Summary of the Invention

[0009] To solve the above technical problem, the present invention provides a method for effectively and correctly classifying interference signals of the same type but unequal distribution, in view of the shortcoming of insufficient generalization in the interference signal classification problem.

[0010] The technical solution adopted by the present invention to solve the above technical problems is a non-identical distribution radar interference identification method based on adversarial transfer learning, comprising the following steps:

[0011] Signal generation and preprocessing steps: First, an ideal interference signal is generated. The ideal interference signal is subjected to time-frequency analysis, smoothing filtering, and adaptive clipping to obtain a source domain sample set consisting of ideal interference signal time-frequency graph samples, and each sample in the source domain sample set is assigned a radar interference type label. Secondly, a target domain sample set consisting of non-identically distributed interference signal time-frequency graph samples is obtained through simulation and addition of non-ideal factors, time-frequency analysis, smoothing filtering, and adaptive clipping.

[0012] The radar jamming recognition steps based on adversarial transfer learning include the source domain pre-learning process, the adversarial discrimination process, and the target domain classification test process:

[0013] Source domain pre-learning process: The labeled source domain sample set is input into the source domain convolutional network. The source domain convolutional network outputs the source domain features to the classifier. The classifier outputs the radar interference classification result and compares it with the label to complete the supervised interference signal recognition training.

[0014] Adversarial discrimination process: unlabeled target domain samples are input into the target domain convolutional network, and unlabeled source domain samples are input into the trained source domain convolutional network. The target domain convolutional network and the source domain convolutional network respectively output the extracted target domain features and source domain features to the target data pre-classifier for adversarial discrimination training. The purpose of adversarial discrimination training is to make it impossible for the target data pre-classifier to distinguish whether the input features are from the source domain or the target domain.

[0015] Target domain classification test process: After inputting unlabeled target domain samples into the target domain convolutional network that has completed adversarial discrimination training for feature extraction, the target domain features are output to the classifier that has completed supervised interference signal recognition training. The classifier completes the final radar interference signal recognition and classification and outputs the final classification recognition result.

[0016] The beneficial effects of the present invention are: 1. It can automatically extract sample features without relying on manual selection; 2. It can achieve both the computational efficiency and generalization of the network and have a high recognition rate; 3. It can effectively classify non-identically distributed samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the processing flow of the embodiment;

[0018] Figure 2 Schematic diagram of signal preprocessing process;

[0019] Figure 3 are examples of source domain and target domain samples;

[0020] Figure 4 This is the ADDA algorithm process. DETAILED DESCRIPTION

[0021] Example method flow as follows Figure 1 As shown in the figure, it includes the signal generation and preprocessing process, the adversarial transfer learning process to output the target domain classification and recognition results, and the final recognition result analysis part.

[0022] 1. Signal Generation and Preprocessing

[0023] Signal preprocessing process such as Figure 2 As shown, including:

[0024] After the ideal interference signal is generated, on the one hand, a source domain sample set consisting of ideal interference signal time-frequency map samples is obtained through time-frequency analysis, smoothing filtering and adaptive clipping. On the other hand, a target domain sample set consisting of non-identically distributed interference signal time-frequency map samples is obtained through simulation and addition of non-ideal factors, time-frequency analysis, smoothing filtering and adaptive clipping.

[0025] Nine types of interference are selected as samples of the ideal interference signal time-frequency diagram, namely, noise amplitude modulation interference AM, noise frequency modulation interference FM, comb spectrum interference COMB, slice forwarding interference CI, intermittent sampling forwarding interference IS, dense copy false target interference SMSP, spectrum dispersion interference MT, convolution modulation smart noise interference CN and suppression deception composite interference. The interference-to-noise ratio generated by the ideal interference is set to four types: 0dB, 2dB, 4dB, and 6dB.

[0026] For the target domain sample set, the non-ideal factor introduced is the presence of strongly suppressive noise combined with environmental conditions. Under complex electromagnetic conditions, strong suppressive noise may exist across the entire receiver frequency band, appearing as a full-screen highlight on the radar display. Adding this noise to the nine ideal interference signals creates a composite interference, another type of non-ideal interference that radar interference detection and identification may encounter in complex electromagnetic environments. In the software implementation, Gaussian white noise is used as the signal simulation for the background strongly suppressive noise. The interference-to-noise ratio (INR) is also used to measure the intensity of this background strongly suppressive composite noise, expressed in dB. The INR generated by the non-identically distributed interference samples is set to -5dB and -10dB.

[0027] The ideal and non-ideal interference signal vectors are normalized, and the time-frequency image is obtained by short-time Fourier transform. The time-frequency image is smoothed by Gaussian filtering to filter out noise as much as possible, and the interference part in the focus frequency band is adaptively cropped. Finally, the time-frequency image is reconstructed and sampled by the nearest neighbor interpolation algorithm, and the image size is unified into a 128*128 bmp format image. The 9 interference sample images in the source domain and target domain sample sets are finally obtained as follows Figure 3 shown.

[0028] (2) Adversarial Transfer Learning Process

[0029] The discriminant model is improved by combining the idea of generative adversarial network (GAN), and the Adversarial Discriminative Domain Adaptation (ADDA) method with non-shared weights is used to identify and classify non-identically distributed radar interference signals.

[0030] The ADDA method can be regarded as a specific adversarial domain adaptation framework, which is trained with the standard GAN loss function. S The training of the classification model f(X) uses the loss function shown in formula (1) Optimize, where K is the number of categories in the sample set to be classified.

[0031]

[0032] X s is the sample space of the source domain, Y s is the source domain label set, x s is the source domain sample to be classified, C is the source domain sample classifier, Indicates that the loss function is calculated for all k types of samples. It means to find the expected value of the distribution specified in the subscript.

[0033] In the adversarial discrimination process, the adversarial discriminator determines whether the sample to be tested is from the source domain or the target domain, and outputs two results: true and false. In this process, both source domain and target domain samples participate in the training. The adversarial discriminator D that distinguishes source domain and target domain samples adopts the loss function shown in formula (2) Optimize.

[0034]

[0035] M T is the feature mapping network of the target domain, D is the adversarial discriminator that distinguishes the source domain and target domain samples, x T is the target domain sample to be classified, Represents the expected value of the distribution of source domain samples in the source domain sample space;

[0036] Finally, corresponding to the training process of the generator in GAN, the loss function shown in formula (3) is adopted The target domain feature mapping network is trained, and the loss function This method shares the same properties as the maximum-minimum loss in GANs, but exhibits a stronger gradient when fitting sample features in the target domain. The maximum-minimum loss is the loss function for generative adversarial GANs, and the training process is a generative adversarial process. The function first seeks the discriminator that maximizes the loss (best discrimination), followed by the generator that minimizes the loss under these conditions. In practice, the two can be trained alternately or simultaneously.

[0037]

[0038] In summary, the ADDA method follows the unconstrained conditions satisfied by equations (1), (2) and (3). Figure 4 As shown in the figure, the ADDA training process is briefly described as first training the source domain samples to obtain a feature mapping network (source domain convolutional network) and a classification model (classifier). Then, using the adversarial idea, the pre-trained source domain feature mapping is used as the initialization target domain representation space, and the source model is modified in the adversarial training to adjust the new target domain mapping network, effectively realizing the feature learning of non-identical distribution samples.

[0039] The ADDA method used shares the parameters of the pre-trained network, reducing the feature loss and the impact on training results caused by insufficient sample quantity and distribution differences between samples. It is an effective way to solve the problems of non-identical distribution between domains and small sample recognition. The ADDA algorithm is divided into three steps: source domain pre-learning process, adversarial discrimination process and target domain classification test process. Figure 4 As shown:

[0040] Source domain pre-learning process: The pre-learning process uses labeled source domain time-frequency graph samples to train the source domain feature mapping network (source domain convolutional network) and the classifier for supervised interference signal recognition. This supervised learning step provides support for the subsequent adversarial discrimination process and the final target domain classification and recognition process. The source domain convolutional network and classifier are the basis for subsequent parameter sharing. After the source domain convolutional network is trained, the unlabeled source domain time-frequency graph samples are input into the classifier to obtain the classification test results.

[0041] Adversarial discrimination process: The adversarial discrimination process uses unlabeled target domain time-frequency graph samples to input into the target domain feature mapping network (target domain convolutional network), and at the same time inputs unlabeled source domain time-frequency graph samples into the source domain feature mapping network (source domain convolutional network) with fixed parameters; the target domain convolutional network and the source domain convolutional network respectively output the extracted target domain features and source domain features to the target data pre-classifier B for adversarial discrimination training. The purpose is to make it impossible for the target data pre-classifier B to distinguish whether the input features are from the source domain or the target domain, so that the feature distribution of the source domain and the target domain is unified, which is manifested by the loss function of the target data pre-classifier B as the adversarial discriminator D and the target domain feature mapping network expressed by formula (2) and formula (3) reaching the maximum and minimum minmax convergence, thereby obtaining the configuration parameters of the target domain convolutional network for target domain feature extraction and sharing the parameters to the target domain convolutional network in the target domain classification test process;

[0042] Target domain classification test process: For the unlabeled target domain interference signal time-frequency map samples, they are input into the target domain convolutional network that has completed adversarial discrimination training for feature extraction. The features are then output to the classifier whose parameters are shared during the pre-learning process. The classifier completes the final radar interference signal recognition and classification and outputs the final classification recognition results. Figure 4 The test process in

[15] also used the source domain convolutional network as a control group to obtain the pre-transfer classification results, see Table 4 for details.

[0043] The structures and parameter designs of the source domain convolutional network, the classifier structure network, and the target data pre-classification network in the embodiment are shown in Tables 1, 2, and 3.

[0044] Table 1 Source domain convolutional network structure design

[0045]

[0046]

[0047] Table 2 Classifier structure network design

[0048]

[0049] Table 3. Target data pre-classifier network structure design

[0050]

[0051] Classification and recognition results

[0052] The experimental simulation platform was Ubuntu 18.04, equipped with an NVIDA GeForce RTX2080Ti GPU. The model was built using the PyTorch framework, and NVIDA Cuda was used to accelerate the computation. In the experiment, the Adam optimizer was used for iteration, with a learning rate of 0.0001, β1 of 0.5, and β2 of 0.999.

[0053] The classification and recognition results of the transferred trained model on the interference signal dataset under conditions of varying sample size and various non-ideal conditions were compared with those of the CNN and the untransferred feature extraction network of the ADDA network. The recognition test results for the target domain sample set are shown in Figure 0. The sample sizes of Sets 1 to 6 increase in ascending order, with Set 1 having the smallest sample size of 540 and Set 6 having the largest sample size of 5400. The recognition rate difference column in the table shows the difference between the post-transfer recognition rate and the pre-transfer recognition rate and the CNN recognition rate.

[0054] Table 4 Results of interference recognition rate in target domain III

[0055]

[0056]

[0057] As can be seen from Table 4, the classification and recognition results of the ADDA structure are significantly improved compared with those before migration and those using the CNN network. The recognition rate of non-identically distributed interference signal sample sets of various sample sizes is generally higher than that of the CNN method.

[0058] The recognition rate after migration of the ADDA method gradually increases with the increase of sample size. The larger the sample size, the more samples are available for training and learning in the adversarial discriminant domain classification process, so the trained feature mapping network can have more comprehensive feature learning and adaptation to the target domain samples. Since the target domain samples used in adversarial domain adaptation are the test classification samples, after adversarial domain adaptation, the target domain feature extraction network has a higher degree of adaptability to the final test classification samples. The recognition rate after migration domain adaptation using the ADDA method is improved by 30-55 percentage points compared to before migration at various sample sizes. Compared with the JR-CNN method, it is improved by 13.9%-27.7% at various sample sizes, and the overall recognition rate reaches 82.6%-99.9%. When the sample size is small, the recognition rate of the ADDA network reaches 82.59%. When the sample size is large, the recognition rate can reach 99.9%;

[0059] For the identification of non-ideal interference in complex electromagnetic environments, adversarial transfer learning improves the generalization of the recognition model, achieving a 13.9%-27.7% improvement in recognition rate compared to CNN methods. When the target domain samples are the test samples and the number of non-identically distributed interference samples is small (only 60 per type of interference), the recognition rate is greater than 80%. The adversarial transfer learning method also achieves excellent recognition of strongly suppressed noise composite interference with a large sample size, reaching over 99% with a sample size of 5400.

Claims

1. A non-identical distribution radar interference identification method based on adversarial transfer learning, characterized by: The following steps are involved: Signal generation and preprocessing steps: First, an ideal interference signal is generated. The ideal interference signal is subjected to time-frequency analysis, smoothing filtering, and adaptive clipping to obtain a source domain sample set consisting of ideal interference signal time-frequency graph samples, and each sample in the source domain sample set is assigned a radar interference type label. Secondly, a target domain sample set consisting of non-identically distributed interference signal time-frequency graph samples is obtained through simulation and addition of non-ideal factors, time-frequency analysis, smoothing filtering, and adaptive clipping. The radar jamming recognition steps based on adversarial transfer learning include the source domain pre-learning process, the adversarial discrimination process, and the target domain classification test process: Source domain pre-learning process: The labeled source domain sample set is input into the source domain convolutional network. The source domain convolutional network outputs the source domain features to the classifier. The classifier outputs the radar interference classification result and compares it with the label to complete the supervised interference signal recognition training. Adversarial discrimination process: unlabeled target domain samples are input into the target domain convolutional network, and unlabeled source domain samples are input into the trained source domain convolutional network. The target domain convolutional network and the source domain convolutional network respectively output the extracted target domain features and source domain features to the target data pre-classifier for adversarial discrimination training. The purpose of adversarial discrimination training is to make it impossible for the target data pre-classifier to distinguish whether the input features are from the source domain or the target domain. Target domain classification test process: After inputting unlabeled target domain samples into the target domain convolutional network that has completed adversarial discrimination training for feature extraction, the target domain features are output to the classifier that has completed supervised interference signal recognition training. The classifier completes the final radar interference signal recognition and classification and outputs the final classification recognition result.

2. The method according to claim 1, wherein: The non-ideal factors are strong suppressive noise and compound environmental conditions.

3. The method according to claim 2, wherein: Gaussian white noise is used as the background for signal simulation of strong suppressive noise compound environment conditions.

4. The method according to claim 1, wherein: Radar jamming types include: noise amplitude modulation jamming, noise frequency modulation jamming, comb spectrum jamming, slice forwarding jamming, intermittent sampling forwarding jamming, dense copy false target jamming, spectrum dispersion jamming, convolution modulation smart noise jamming and suppression deception composite jamming.

5. The method according to claim 1, wherein: The interference-to-noise ratios (INRs) of ideal interference in the signal generation and preprocessing steps are set to 0dB, 2dB, 4dB, and 6dB; the INRs of the time-frequency graph samples of non-identical interference signals are set to -5dB and -10dB.

Images