Research On Small Sample Radar Signal Recognition Based On Unsupervised Deep Clustering

The combination of cluster analysis method based on deep learning technology and individual recognition of radar signals is a direction worth studying.Deep clustering models often require a large amount of training data,and it is very common that sparse annotation data in actual industrial scenarios or have unbalanced data samples.Based on the problem of few samples,this paper designs a new Inverse autoencoder(IAE)for radar signals from different radiation sources collected in a complex external environment,and combines it with the Generative Adversarial Network(GAN).A new unsupervised deep clustering model IAE-Cluster GAN is proposed.The main work and contributions of the paper are as follows:Firstly,A new unsupervised deep clustering model IAE-Cluster GAN is proposed,which can control the distribution type of the learned latent codes without additional constraints,thereby ensuring that the clustering structure and semantic information of the original data can be more completely retained in the learned latent space.In addition,the attention mechanism is integrated into the network model,so that the latent code contains more useful clustering information;the use of hyperspherical mapping improves the stability of model training and reduces the training parameters.Secondly,It is proposed to use a generative countermeasure network to generate a bispectral matrix similar to the original data sample to solve the problem of scarcity and imbalance in the number of samples in radar signal data samples.Use hyperspherical mapping to replace the gradient penalty term in the original loss function,and improve the network without introducing new parameters;use Spectral normalization in the discriminator network to ensure that the discriminator network is K-Lipschitz is continuous,which improves the stability of network training.Thirdly,The unsupervised deep clustering model IAE-Cluster GAN proposed in this paper is applied to the small sample radar signal dataset.And introduce contrast loss in the objective function to solve the problem of high overlap of samples of different categories in the representation space at the beginning of the clustering task,so that the spacing between clusters in the learned embedding space is more obvious and the clustering accuracy is improved.