Radar working mode amplitude waveform augmentation based on generative adversarial networks

In radar working mode recognition, one of the most widely utilized data formats is working mode data in the form of amplitude waveform sequences. However, the capacity to train big, sophisticated deep learning models is somewhat limited because this type of data involves military secrets and is costly to gather and annotate. This research suggests a generative adversarial network-based radar working mode amplitude waveform augmentation technique to address this issue. Firstly, the technique aims to amplify the radar working mode data set by artificially synthesizing limited amplitude waveform data using generative adversarial networks (GANs) and their variations: Wasserstein GANs (WGANs), Auto-encoders GANs (AE-GANs), and Variational Auto-encoders GANs (VAE-GANs). Secondly, this research suggests a type of Cross-domain Generative Adversarial Networks (CD-GANs) that use amplitude and waveform sequence data to generate each other by utilizing the possible mapping link between amplitude and waveform. To achieve data augmentation, the network can be trained on pre-existing datasets to produce class-specific synthetic data of any length, extending or substituting real amplitude waveform sequences of the same mode of working. Then, to assess the quality of the synthetic data produced by the aforementioned model, the qualitative visual assessment of the synthetic data was conducted using the visualization evaluation metrics of principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE), while the quantitative evaluation metrics of wavelet coherence scores were used to compare the similarity between the synthetic and real data. Finally, by training the classifier utilizing a hybrid strategy of synthetic and real data using a standard Convolutional Neural Networks (CNNs) classification model, the effectiveness of synthetic data in the training phase of the radar working pattern recognition model is confirmed. In summary, the simulation experiment confirms the superiority of the CD-GANs model and shows that its generated data is more accurate than that of other GANs models. It can also achieve more realistic and trustworthy augmentation of radar amplitude waveform data. ● The GANs and their variations (WGANs, AE-GANs, VAE-GANs) are employed instead of the conventional PDW level data augmentation to enhance the radar radiation source working mode amplitude sequences. ● The potential mapping relationship between amplitude and waveform sequences in the same operating mode is used to suggest a Cross-domain Generative Adversarial Networks (CD-GANs) model where amplitude and waveform sequences mutually create each other. It can be used to expand or replace the actual data of the same working mode, generate data of any length and in any category, and accomplish data augmentation. ● The quality of the synthetic data generated by the model was evaluated using both qualitative and quantitative techniques. The visual evaluation was conducted using t-SNE and PCA, two qualitative evaluation indices. The wavelet coherence score, a quantitative evaluation metric, was employed to compare the degree of resemblance between the synthesized and genuine data. ● The effectiveness of synthetic data in the training phase of the radar working mode recognition model is validated, and the superiority of the amplitude waveform augmentation model based on CD-GANs is further validated, by training classifiers using a hybrid strategy of synthetic and real data using a standard CNN classification model.