Scent classification by K nearest neighbors using ion-mobility spectrometry measurements

• Published in: Expert systems with applications 2019-01, Vol.115, p.593-606
• Main Authors: Müller, Philipp, Salminen, Katri, Nieminen, Ville, Kontunen, Anton, Karjalainen, Markus, Isokoski, Poika, Rantala, Jussi, Savia, Mariaana, Väliaho, Jari, Kallio, Pasi, Lekkala, Jukka, Surakka, Veikko
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Classification; Classifiers; Data analysis; Electronic noses; Ion-mobility spectrometry; Ionic mobility; K nearest neighbours; Machine learning; Principal components analysis; Retraining; Scent classification; Scientific imaging; Searching; Spectrometry; Spectroscopy
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2018.08.042

•K Nearest Neighbor classifies scents and their chemical components.•Scents/ chemicals are classified using only ion-mobility spectrometry measurements.•Classification using k-dimensional tree search is approximately 8-times faster.•By principal component analysis 71–86% of features are ignored for classification. Various classifiers for scent classification based on measurements using an electronic nose (eNose) have been studied recently. In general, classifiers rely on a static database containing reference eNose measurements for known scents. However, most of these approaches require retraining of the classifier every time a new scent needs to be added to the training database. In this paper, the potential of a K nearest neighbors (KNN) classifier is investigated to avoid the time-consuming retraining when updating the database. To speed up classification, a k-dimensional tree search in the KNN classifier and principal component analysis (PCA) are studied. The tests with scents presented to an eNose based on ion-mobility spectrometry (IMS) show that the KNN method classifies scents with high accuracy. Using a k-dimensional tree search instead of an exhaustive search has no significant influence on the misclassification rate but reduces the classification time considerably. The use of PCA-transformed data results in a higher misclassification rate than the use of IMS data when only the first principal components explaining 95% of the total variance are used but in a similar misclassification rate when the first principal components explaining 99% of the total variance are used. In conclusion, the proposed method can be recommended for classifying scents measured with IMS-based eNoses.