Detection of outliers in surface acoustic wave (SAW) chemical sensor array responses by one-class support vector machine

Abstract

Usefulness of one-class support vector machine (SVM) is demonstrated for detection of outliers in surface acoustic wave (SAW) sensor array data for odor recognition. The one-class SVM for outlier detection is essentially a two-class pattern recognition formulation wherein outliers are considered to be the only target class, and the rest data points are grouped to make a normal class. The construction of decision function needs training with known class identities. In test phase, the algorithm picks up those data points as outlier which do not classify as normal. The SAW sensor array is an important platform for making electronic noses that fulfill varied needs of specific applications. The outliers in SAW electronic noses may occur due to electronic instabilities, fluidic fluctuations, electromagnetic interference, temperature fluctuation, or presence of non targeted chemicals. Therefore, it is important to clean up data for outliers for achieving high performance SAW odor recognition system. Even though one-class SVM has been used in other applications such as image processing and text recognition, here we analyze its suitability for sensor array based electronic nose data. A simulated SAW sensor array dataset comprised of 11 sensors and 6 vapor classes laden with varied levels of noise and outliers is considered. It is shown that the one-class SVM is quite efficient for detection of outliers in electronic nose data.