Background signal estimation from multi-sensor signals based on outer product and non-linear filters

The source separation based on a statistical and computational technique is one of the most exciting topic in a multivariate signal analysis. We proposed a novel source separation technique using a tensor product expansion. This technique is the signal separation that the background noise, which is observed in almost all input signals, can be estimated by using a tensor product expansion where the absolute error is used as the error function (TPE-AE). The effectiveness of TPE-AE for artificial signals and real signals is represented in [1]. However, TPE-AE has two problems; one is that a result of TPE-AE is strongly affected by Gaussian random noise, another one is that an estimated signal varies widely due to a random search. In order to solve these problems, signal estimation techniques based on a median and Modified Trimmed Mean (MTM) are proposed in this paper. These methods calculate the outer product using a non-linear filter to estimate the background noise. Results show that novel techniques can separate the background noise more accurately than the conventional TPE-AE.