Application of independent component analysis to detection of gas leakage sound

It is important to detect the leakage of the gas to be flammable or poisonous from cracks in pipes of chemical plants. We use sound to detect the gas leakage. It is necessary to examine the proper feature extraction for the sound to get the high detection performance. We applied independent component analysis (ICA) to feature extraction. The purpose of this study is to evaluate the effectiveness of feature extraction using ICA. Experiments were performed in a plant using an artificial gas leakage device under various experimental conditions. We collected leakage sound and background noise around a noisy machine. Most of the basis functions trained with the collected acoustic signal were localized in frequency. Furthermore, there were remarkable differences in amplitudes of some independent components between the leakage sound and the background noise. These results indicate that the ICA was effective for the feature extraction of the leakage sound