Machine learning techniques to select a reduced and optimal set of sensors for the design of Ad Hoc sensory systems

The first step of this research has been to discriminate, by means of a commercial electronic nose (e-nose), the maturity evolution of seven types of fruits stored in refrigerated cells, from the post-harvest period till the beginning of the marcescence. The final aim was to determine a procedure to select a reduced set of sensors that can be efficiently used to monitor the same class of fruits by a low cost system with few, suitable sensors without loss in accuracy and generalization. To define the best subset we have compared the use of a projection technique (the Principal Component Analysis, PCA) with the sequential feature selection technique (Sequential Forward Selection, SFS) and the Genetic Algorithm (GA) technique by using classification schemes like Linear Discriminant Analysis (LDA) and k-Nearest Neighbor (kNN) and applying two data pre-processing methods. We have determined a subset of only three sensors which gives a classification accuracy near 100%. This procedure can be generalized to other experimental situations to select a minimal and optimal set of sensors to be used in consumer applications for the design of ad hoc sensory systems.