Feature extraction for gas photoacoustic spectroscopy and content inverse based on overcomplete ICA bases

In near-infrared band, macro representations of rovibrational-transition absorptions of some low molecule mass gases are “spiky” transient lines. The pressure broadening at ambient temperature and pressure make lineshape profiles at adjacent wavelength produce mutual projection or superposition, and it makes the “spiky” absorption features just be hidden in observable spectra. A blind source separation (BSS) model can extract features which hide within sampled spectra data, if a mixed gas phase photoacoustic signal is regarded as the weighted sum of feature absorptions by the spacing between adjacent feature wavelengths. A BSS model based on overcomplete ICA basis is proposed, and a weight truncation equation for transforming ICA bases based on the five-point sampling method is created so that a FastICA algorithm based on a negentropy approximation can be used to extract feature components. The resulting validity test using data from a real-time experiment of gas detection showed that the method detection limit can be improved from 16 ppb to 10 ppb with improved accuracy and signal-to-noise ratio.

► A overcomplete-ICA-basis model.
► A truncation equation.
► Five-point sampling method.