Prediction of human assessments of odor using electronic nose and neural networks

Use of instruments instead of human panels to assess odors can save money and make the collection process more efficient. In this work, an electronic nose comprising of an array of 32 internal polymer sensors was applied to collect measurements from dairy farm odor sources. Artificial neural networks using the Levenberg-Maerquardt Back-propagation algorithm were used to build prediction models to predict human response to odor pleasantness on a generic hedonic scale that ranged from -11 (extremely unpleasant) to +11 (extremely pleasant). Forward selection (FS), Gamma Test (GT), and Principal Component Analysis (PCA) were used to reduce the noise of the measurements. For 28 input candidates, eight variables were selected using GT, and nine variables were selected using FS and PCA, respectively. GT-LMBNN was considered as the highest accurate model based on the probability of separate validation results falling within the human assessments range.