AN APPROACH TO CORRELATION-BASED SOURCE SEPARATION

摘要

This paper reveals a simple, but substantially noise-resistant MIMO decomposition algorithm with very low computational complexity. It is suitable for convolutive signal mixtures induced by binary signal sources. The algorithm eliminates the influence of system channel responses from the system output observations by implementing an inverse of the output correlation matrix. This leads to satis-fiable separation of sources even if they are not completely orthogonal, if their number is underestimated, and if the measurements are noisy. A separation index is defined which points out the time instants of a source activation. We exemplify the algorithm's operation and performance by surface electromyogram (SEMG) decomposition. A six-electrode measurement on five SEMG sources was simulated. In spite of a too low number of measurements to guarantee a thorough source sepatarion, the decomposition results show 100% accuracy in detection of innervation pulses, even with 0 dB additive noise.