Diagonally Loaded Normalised Sample Matrix Inversion (LNSMI) for Outlier-Resistant Adaptive Filtering

Abstract

Instead of a "hard" decision on ignoring "outlier" training samples in constructing the covariance matrix estimate, we propose a "softer" method that reduces the impact of such abnormal data samples on adaptive filter performance. Specifically, we introduce a diagonally loaded covariance matrix estimate that is normalised by a generalised inner product (GIP), which is more robust against outliers. We demonstrate the efficiency of this technique on high-frequency (HF) over-the-horizon radar (OTHR) data.