In this paper, experiments on the application of the independent component analysis (ICA) technique to separate unknown source signals are reported. ICA is one of the fastest growing fields in signal processing with applications to speech recognition systems, telecommunications, and biomedical signal processing. It is a data-transformation technique that finds independent sources of activity from linear mixtures of unknown independent sources. The statistical method to measure independence is to find a linear representation of the non-Gaussian data so that the components are as independent as possible and the mutual information between them is minimum. Although extensive simulations have been performed to demonstrate the power of the learning algorithm for the problems of instantaneous mixing and un-mixing of sources, its application to the noise diagnosis and separation in an industrial setting has not been considered. Noise separation in machinery has a strong basis in the “cocktail problem” in which it is difficult to separate/isolate the voice of a person in a room filled with competing voices and noises. The experiments conducted consist of separating several artificially generated sources of noise. Our results demonstrate that ICA can be effectively employed for such kinds of applications. The underdetermined problem in which there are fewer sensors than sources in the ICA formulation is also examined by applying a time-invariant linear transformation of the acquired signals to identify a single source.
A new proof of a contrast function for bounded component analysis and further analysis
