ELECTRONIC NOSE AND PROBABILISTIC NEURAL NETWORK USE FOR SAUSAGES IDENTIFICATION

UDC 637.07:637.068

Abstract

The electronic nose system based on seven quartz-microbalance sensors was used to generate a pattern of the volatile compounds present in sausage samples with different mass content of soy protein isolate (SPI) ranging from 0 to 30% w/w. The original response curve was extracted to two features such as the maximum response value (ΔF_i^max, Hz) and area values of sensor response curve and time axis surrounded (S_i, Hz×s). All parameters subjected to pattern recognition analysis in original, normalised, scaling values and after principal component analysis. In this paper we used probabilistic neural network (PNN) for multiclass discrimination of sausage products. The neural network architecture was optimized for samples discrimination using as input vectors the electronic nose parameters such as S_i and ΔF_i^max. The best classification reliability (95,8%) for model based on dataset of S_i in original values obtained with the values of the PNN smoothing parameter σ in the range from 3,6 to 54,0. The classification model built by ΔF_i^max dataset in original values gave the 100% identification accuracy with the value of the PNN smoothing parameter in the range from 0,2 to 1,0. Furthermore, the different pre-processing techniques applied to datasets led to a slight decrease the prediction performance of the classification models, but the speed of neural network training has increased.

This paper presents a novel approach to identification of cooked sausages and determination of soy products mass content using the electronic nose combined with probabilistic neural network. Compared to classical methods, this new approach could represent an alternative and innovative tool for faster and cheaper sausage identification and mass content of soy protein isolate (0, 10, 20, 30% w/w) detection.