Classification of microbial defects in milk using a dynamic headspace gas chromatograph and computer-aided data processing .2. Artificial neural networks, partial least-squares regression analysis, and principal component regression analysis

Objective, yet cost-effective evaluation of flavor is difficult in quality control of milk. Inexpensive gas chromatographs in conjunction with computer models make it feasible to construct an objective flavor evaluation system far routine quality control purposes. The purpose of this study was to classify milk with microbial off-flavors using a low-cost headspace gas chromatograph and computer-aided data processing. Principal component similarity (PCS) analysis was discussed in part 1. In part 2, artificial neural networks (ANN), partial least-squares regression (PLS) analysis, and principal component regression (PCR) analysis are examined. UHT milk was inoculated with various bacteria (Pseudomonas fragi, Pseudomonas fluorescens, Lactococcus lactis, Enterobactor aerogenes, and Bacillus subtilis) and a mixed culture (P. fragi:E. aerogenes:L. lactis = 1:1:1) to approximately 4.0 log(10) CFU mL(-1). ANN were able to make better predictions than PLS and PCR. The prediction ability of PLS was better than PCR. The performance of each method depended on the content of training and testing of data, i.e., more data resulted in better predictive ability.