Self-evolving Radar Target Detection Using Deep Learning Methods

Radar target detection technology is widely used in military and civil aviation activities.With the development of stealth technology and the increase of flying speed,radar targets have features such as high clutter,low signal-to-noise ratio,and excessively fast movement speeds.Higher requirements for radar target detection technology are proposed.Deep learning has strong capability for automatic extraction,which can avoid the limitation of the traditional methods to extract features manually,and has achieved certain results in the field of radar target detection.Due to weather,radar hardware aging and other factors,radar data may change as time passes.The initial training deep learning model may not be applicable to subsequent data,so new data need to be used to retrain the model so that the model can evolve as data changes.The evolution of the model requires a large amount of labeled samples to train the model.Although the radar data is easy to obtain,the labeling of the data tag is very costly.How to use the unlabeled data to evolve the deep learning model is very important.For the problems mentioned above,this paper proposes a self-evolving radar target detection algorithm using deep learning methods,which uses an unsupervised approach to achieve self-evolution of the deep learning model.Firstly,two radar detection models based on deep learning are designed in this paper.These two modes are deep belief networks(DBN)and recurrent neural networks(RNN).They use supervised training to generate the basic model of radar target detection.Secondly,based on these two models,the unsupervised self-evolution of deep learning model is realized using the dual-view cooperative training algorithm.The key of the dual-view cooperative training algorithm lies in the construction of dual views.This paper constructs dual views through two different directions.They are dual-view model based on single-period and multi-period feature construction and dual-view model based on DBN and RNN models.In addition,this paper also presents a self-evolving radar target detection algorithm that combines dual-view cooperative training with active learning.In active learning,the model actively selects some samples according to the learned knowledge and submits it to experts to label labels,which are then used retrain the model.Active learning improves the problem of passively accepting samples in dual-view collaborative training.Finally,the three-day data collected from a radar is used to verify this algorithm.The results show that the proposed algorithm can achieve the self-evolution of the radar target detection model through an unsupervised or semi-supervised approach.