Estimation of SNR for modulation recognition in the presence of HF noise

Abstract

The subject of automatic modulation recognition has been studied for decades, but recently has received much more attention with the capabilities of modern computing systems and software defined radio (SDR) platforms. The two most common modulation recognition methods are feature-based (FB) and likelihood-based (LB). Researchers of both types develop models and algorithms around assumptions of additive white Gaussian noise and often further assume an ability to estimate the signal-to-noise ratio (SNR) with some confidence. Recent investigations [1]–[3] have reconfirmed that in the high-frequency (HF) band, the assumption of Gaussian noise is not always valid, since noise in this band is strongly affected by impulsive atmospheric radiators, such as lightning. It appears that a modified Bi-Kappa distribution is better suited for modelling such noise. And, since the estimation of SNR is an important factor in modulation recognition algorithms, it is prudent to investigate the effects that non-Gaussian noise has on SNR estimators. One moments-based estimator [2] is Aisbett's [4] hash function and shows promise as an accurate estimator over a large range of signal-to-noise ratios under the assumption of Gaussian noise. This paper extends the earlier work [2] and investigates the performance of the SNR estimator in the presence of Bi-Kappa distributed noise with initial results suggesting that Bi-Kappa distributed noise degrades the performance of the SNR estimator.