Intrinsic Low-Dimensional Nonlinear Manifold Structure of Radio Frequency Signals

Abstract

Nonlinear dynamic features are crucial in specific emitter identification (SEI), which works based on the nonlinearity of the emitter. Nevertheless, its underlying principles and stability have not been systematically studied, so these methods are close to a black box, making the recognition performance of the system unstable in applications. In this study, we provide an in-depth analysis of its mechanism and theoretical model, extracting, for the first time, the intrinsic low-dimensional nonlinear manifold structure (IMS) of radio frequency signals based on manifold learning and analyzing its robustness and distinctiveness for SEI. Hardware type and intentional modulation are found to determine the overall shape of IMS, and the device-specific information for SEI is subtle and distributed locally on IMS. Notably, sampling-related random disturbances cause the distortion, zooming, or rotation of IMS, which lowers the identification accuracy by a range of 1.74%–52%, depending on the feature, device type, signal type, and other factors.