Classification models and algorithms in application of multi-sensor systems to detection and identification of gases

Abstract

The objective of the thesis is to adopt advanced machine learning tech- niques in the analysis of the output of sensor systems. In particular we have focused on the SVM (Support Vector Machine) approach to classi- ¯cation and regression, and we have tailored such approach for the area of sensor systems of the "electronic nose" type. We designed an Electronic Nose (ENose), containing 8 sensors, 5 of them being gas sensors, and the other 3 being a Temperature, a Humidity, and a Pressure sensor, respectively. Our system (Electronic Nose) has the ability to identify the type of gas, and then to estimate its concentration. To identify the type of gas we used as classi¯cation and regression technique the so called Support Vector Machine (SVM) approach, which is based on statistical learning theory and has been proposed in the broad learning ¯eld. The Kernel methods are applied in the context of SVM, to improve the classi¯cation quality. Classi¯cation means ¯nding the best divider (separator) between two or more di®erent classes without or with minimum number of errors. Many methods for pattern recognition or classi¯cation are based on neural network or other complex mathematical models. In this thesis we describe the hardware equipment which has been designed and implemented. We survey the SVM approach for machine learning and report on our experimentation.