Encoding FT-IR Spectra in a Hopfield Network and Its Application to Compound Identification in Open-Path FT-IR Measurements

Vapor-phase Fourier transform infrared (FT-IR) reference spectra of five alcohols (methanol, ethanol, 1-propanol, 2-propanol, and 1-butanol) from 1277 to 891 cm - 1 were encoded as prototype vectors for a Hopfield network, along with four baseline spectra in the same spectral region. For each input spectrum, the Hopfield network is designed to identify one component of a mixture at a time (usually the major component). Inclusion of the baseline spectra in the Hopfield network allowed the network to classify spectra as unknowns, when they were not stored as prototype vectors in the network. A method that can minimize the number of spurious points and maximize the domain of attraction of the designed equilibrium points was used to design the Hopfield network. Over 100 spectra of different compounds were input to this network, and only three of them were incorrectly classified. The network could usually identify each of the five alcohols correctly even in the presence of noise and interfering compounds. When random noise was added to each spectrum up to a level of 20% of the maximum absorbance of the analyte, the network could still identify each alcohol with 100% accuracy. The network could also correctly identify one component of representative mixture spectra. The Hopfield network was successfully used to identify ethanol and 1-butanol from their open-path FT-IR (OP/FT-IR) spectra despite the large degree of similarity of these spectra, the weakness of the bands, and the fact that the measured spectra were dominated by lines due to the absorption of atmospheric water vapor.