Classification and Authentication of Mineral Water Samples using Electronic Tongue and Deep Neural Networks

Supervised multiclass classifiers based on deep neural networks (one-dimensional convolutional neural network (1D-CNN) and long short-term network (LSTM)) are developed to classify and authenticate mineral water samples of commercial brands (Aquafina, Bisleri, Oasis, Kingfisher, Dolphin and McDowell) available in Indian market. Electronic tongue based experiments are conducted to generate output waveforms (current signals) of the water samples of the six brands. Since the data obtained via the experiments are not adequate to train the deep neural networks, Variational Autoencoder (VAE) is used to generate additional current signals for each of the six brands. New database consisting of the experimental data and data generated using VAE is utilized to train and test the classifiers. Classification accuracies obtained by the LSTM and 1D-CNN models are 83.33% and 66.66%, respectively.