Modulation Classification of MIMO-OFDM Signals by Independent Component Analysis and Support Vector Machines

TL;DR

This paper presents a blind modulation classification scheme for MIMO-OFDM signals using ICA for signal separation and SVM or ML for classification, effective over frequency selective, time-varying channels.

Contribution

It introduces a novel combination of ICA and machine learning for blind modulation classification in MIMO-OFDM systems, handling channel variations.

Findings

High probability of correct classification over realistic channels

Performance improved by exploiting channel invariance

Upper bound on classification accuracy derived

Abstract

A modulation classification (MC) scheme based on Independent Component Analysis (ICA) in conjunction with either maximum likelihood (ML) or Support Vector Machines (SVM) is proposed for MIMO-OFDM signals over frequency selective, time varying channels. The method is blind in the sense that it is assumed that the receiver has no information about the channel and transmitted signals other than that the spatial streams of signals are statistically independent. The processing consists of separation of the MIMO streams followed by modulation classification of the separated signals. While in general, blind separation of signals over frequency selective channels is a difficult problem, the non-frequency selective nature of the channel experienced by individual symbols in a MIMO-OFDM system enables the application of well-known ICA algorithms. Modulation classification is implemented by maximum likelihood and by an SVM-based modulation classification method relying on pre-selected modulation-dependent features. To improve performance in time varying channels, the invariance of the channel is exploited across the coherence bandwidth and the time coherence. The proposed method is shown to perform with high probability of correct classification over realistic ITU pedestrian and vehicular channels. An upper bound on the probability of correct classification is developed based on the Cramer Rao bound of channel estimation.