2003 Volume 16 Issue 10 Pages 539-547
It is important to detect gas leakage sounds from pipes in petroleum refining plants and chemical plants, as often the gas used in these plants are flammable or poisonous. In order to detect the leakage accurately, we should select a feature extraction method for sounds properly. The purpose of this paper is to examine whether independent component analysis (ICA) is useful as a feature extraction method. Several experiments are performed in a plant using an artificial gas leakage device under various experimental conditions. A separating matrix that separates the independent components from collected leakage sounds and background noises is trained by an ICA algorithm. Through several simulations, we find that most basis functions acquired from this training are localized in frequency. Furthermore, there are remarkable differences in amplitude of some independent components between leakage sounds and background noises. From these results, we confirm that the feature extraction using the ICA algorithm is very useful for detecting gas leakage sounds.