Application of an Artificial Neural Network to the Identification of Amino Acids from near Infrared Spectral Data

Abstract

An artificial neural network (ANN) was trained to identify a group of amino acids from near infrared (NIR) spectral data. The input, the hidden and the output layers were composed of 701 (for raw spectral data, fixed) or 324 (for second derivative of the spectral data, fixed) units, 1 to 100 (changeable) units and 20 (fixed) units, respectively. Using the raw spectral data, the ANN did not converge to a suitable error level. However, when the second derivative spectra were used, whether original or standardised spectra, the error reduced to a suitable level, because this mathematical treatment made their differences in NIR spectra clearer. The ANN was trained for non-pretreated amino acids and then applied to the other prediction sets. When standardised spectra were used, the ANN could almost correctly identify the amino acids not only for non-pretreated amino acids but also for ground samples or samples from different batches. The results obtained by principal component analysis (PCA) were also compared with those by the ANN.