Recognition of Unknown Radar Emitter Based on USRP and Machine Learning

Abstract

Aiming at the problem that the existing priori recognition methods basically fail when identifying newly emerging unknown radar emitter, a method for identifying unknown radar emitter based on machine learning is proposed. First, the designed convolutional neural network can be used to identify four known signals. Based on the transfer learning method, the trained and untrained signals are input into the CNN respectively, and 256 neurons in the fully connected layer are extracted as the features of the known signal and the unknown signal. Then, based on signal characteristics and actual application requirements, random forest is used to identify unknown radar emitter signals. Under the condition of -6dB 9dB signal-to-noise ratio, the overall recognition rate of unknown signals can reach 72.6%. By changing the types of unknown signals and increasing the unknown, the method has proved to have good stability further. Finally, signal generators and software radio peripherals are used to build a signal transceiver experimental system, and a signal recognition interactive interface has been designed, the recognition network is optimized and fine-tune by the actual signal received. The recognition method proposed in this article is verified by adding a new unknown signal further.