Multi-Window Feature Extraction for SVM-Based Electronic Nose Classification of Four Herbal Essential Oils Bambang Heru Iswanto, Fajar Hardoyono, Haris Suhenda, Mutia Delina
Department of Physics, Universitas Negeri Jakarta, Jl. Rawamangun Muka, Jakarta 13120, Indonesia
Abstract
Rapid authentication of herbal essential oils requires sensor models that are accurate, fast, and chemically interpretable. This paper presents an electronic-nose workflow for four oils, namely red ginger, white turmeric, turmeric, and lemongrass, using multi-window feature extraction and support vector machines. Signals were collected from a ten-sensor metal-oxide-semiconductor array during a 10-70 s exposure period after baseline correction. Features were extracted from short, medium, and long windows corresponding to onset, peak development, and tail dynamics, then classified using radial-basis-function SVMs under nested cross-validation. GC-MS profiles were used as class-level chemical anchors rather than direct regression targets. Early windows already contained strong class information, especially for citral-rich lemongrass, whereas mid and late windows improved interpretation for slower terpene-rich oils. The compact window triad produced competitive macro-F1 performance while preserving an auditable link between sensor features and volatile kinetics. The workflow supports rapid screening of herbal essential oils with a parsimonious and chemically interpretable feature set.
