Multi-Channel missing data recovery by exploiting the low-rank hankel structures

Abstract

This paper studies the low-rank matrix completion problem by exploiting the temporal correlations in the data. Connecting low-rank matrices with dynamical systems such as power systems, we propose a new model, termed multi-channel low-rank Hankel matrices, to characterize the intrinsic low-dimensional structures in a collection of time series. An accelerated multi-channel fast iterative hard thresholding (AM-FIHT) with a linear convergence rate is proposed to recover the missing points. The required number of observed entries for successful recovery is significantly reduced from conventional low-rank completion methods. Numerical experiments are carried out on recorded PMU data to verify the proposed method.