Noise resilience through band-limitation in signal regression analysis

Linear regression is used in signal analysis when other methods like artificial neural networks or support vector machines either lack the ability to represent the result in form of a signal or cannot be applied to continuous target values. However, signal noise may lead to unstable noisy solutions with bad performance on non-trained data, especially for underdetermined systems. This work develops a method to add statistical virtual noise with special properties such as band-limitation to the signals in order to reduce these properties in the solution signal. The results show stable solutions with significantly improved performance on non-trained data. The method is also tested on real EEG data.