Characterization and pattern recognition of selected sensors for food safety applications

This study has developed and evaluated the regioregular poly (3-hexyl thiophene) (rr-P3HT) based chemoresistive and piezoelectric sensors to detect the alcohol volatile organic compounds (VOCs) found to be present in the spoiled and Salmonella typhimurium contaminated beef package headspace gas. Development of robust pattern recognition algorithms was another major component of the study. This study evaluated beef experiment datasets related to two types of sensor systems (custom built in our laboratory); thin-film (TF)-module electronic nose system and integrated sensor system. Adaptive wavelet packet transform (WPT) based feature extraction techniques were used for the classification of contaminated packaged beef from the uncontaminated ones. Moreover, the issue of small datasets in artificial neural network (ANN) based beef classification has been investigated by implementing the synthetic sample generation techniques.;The rr-P3HT based chemoresistive sensors were developed using two types of dip coating methods; vertical dip (90°) and inclined dip (10°) coating. The sensors developed using vertical and inclined dip coating method were found to provide repeatable, reproducible, and selective response to trace level concentrations of 3-methy1-1-butanol and 1-hexanol, with the lower detection limit (LDL) of 10 parts per million (ppm) and 12 ppm, respectively. The piezoelectric polymer sensors developed using drop coating technique were found to provide repeatable, reproducible, and selective response to trace level concentrations of 3-methyl-1-butanol and to hexanol with the LDL of 4 ppm and 3 ppm, respectively.;The pattern recognition research involved implementation of WPT based feature extraction techniques on packaged beef (uncontaminated and Salmonella inoculated). The performance of adaptive wavelet transform based feature extraction algorithms was compared with the standard wavelet transforms based feature extraction algorithm. This study also evaluated mega-trend diffusion (MTD) and functional virtual population (FVP) techniques of data domain expansion along with multivariate normal (MVN) synthetic sample generation on small datasets associated with packaged beef contamination detection. The average overall packaged beef classification accuracies of the six synthetic datasets (generated from the corresponding original beef experiment acquired using integrated sensor system) were in the range of 86.7% to 98.9%.