Application of manifold learning algorithms to improve the classification performance of an electronic nose

Abstract

The problem of classification and the use of robust methodologies in electronic nose sensor arrays applications are still open. Among the several steps in the developed method- ologies, it is clear that feature extraction methods are one of the tools that allow improving the classification accuracy in this kind of sensors. As a contribution to solving this problem, in this work, a machine learning methodology is introduced to improve the signal processing and the development of methodologies in the classification task with electronic noses. The methodology considers a normalization to scale the data from the sensors, and four manifold learning algorithms were compared to reduce the dimensionality of the data at the input of the classification machine, these methods are Kernel PCA, Isomap, Laplacian Eigenmaps, and Locally Linear Embedding. The classifier used as a machine learning algorithm was K-NearestNeighbor (KNN) with Euclidean distance and one neighbor. To validate the methodology, a dataset of 3600 measurements that measured six volatile organic compounds is used. Holdout cross-validation approach was used to evaluate the proposed methodology, results showed the best classification accuracy of 98.33% was reached by Kernel PCA as a nonlinear feature extraction method. Finally, the influence in the number of dimensions used to reduce the data was determined to establish the correct number according to the best classification rate in the classification process.