- Original Pdf: pdf
- Keywords: Electronic Nose, EVA, modular, olfaction, sensitivity, selectivity, analyte, temperature oscillated waveforms, features, fingerprint
- TL;DR: On this paper we will discuss the process of sniffing volatile organic compounds from liquid beer samples and exploring various machine learning models
- Abstract: Olfaction has been and still is an area which is challenging to the research community. Like other senses of the body, there has been a push to replicate the sense of smell to aid in identifying odorous compounds in the form of an electronic nose. At IBM, our team (Cogniscent) has designed a modular sensor board platform based on the artificial olfaction concept we called EVA (Electronic Volatile Analyzer). EVA is an IoT electronic nose device that aims to reproduce olfaction in living begins by integrating an array of partially specific and uniquely selective smell recognition sensors which are directly exposed to the target chemical analyte or the environment. We are exploring a new technique called temperature-controlled oscillation, which gives us virtual array of sensors to represent our signals/ fingerprint. In our study, we run experiments on identifying different types of beers using EVA. In order to successfully carry this classification task, the entire process starting from preparation of samples, having a consistent protocol of data collection in place all the way to providing the data to be analyzed and input to a machine learning model is very important. On this paper, we will discuss the process of sniffing volatile organic compounds from liquid beer samples and successfully classifying different brands of beers as a pilot test. We researched on different machine learning models in order to get the best classification accuracy for our Beer samples. The best classification accuracy is achieved by using a multi-level perceptron (MLP) artificial neural network (ANN) model, classification of three different brands of beers after splitting one-week data to a training and testing set yielded an accuracy of 97.334. While using separate weeks of data for training and testing set the model yielded an accuracy of 67.812, this is because of drift playing a role in the overall classification process. Using Random forest, the classification accuracy achieved by the model is 0.923. And Decision Tree achieved 0.911.