Perfume Classification Using a Low-Cost Electronic Nose with Isolation Forest and Linear Discriminant Analysis

Abstract

This research proposes a perfume classification methodology based on a low-cost electronic nose (e-nose) system utilizing MQ gas sensors. The proposed methodology incorporates a preprocessing pipeline featuring Z-score standardization for data normalization, Isolation Forest for efficient outlier detection, Linear Discriminant Analysis (LDA) for effective feature extraction, and Radial Basis Function Support Vector Machine (RBF-SVM) for classification. Experimental evaluations conducted on ten distinct perfume samples using six MQ sensors (MQ-2, MQ-3, MQ-4, MQ-5, MQ-7, and MQ-135) demonstrate that the LDA-based approach achieves a classification accuracy of 95%, outperforming other dimensionality reduction methods like Principal Component Analysis (PCA) at 85%, Uniform Manifold Approximation and Projection (UMAP) at $75 \%$, and t-Distributed Stochastic Neighbor Embedding (t-SNE) at 85%. The robustness of LDA approach is further validated through 5-fold cross-validation, yielding a mean accuracy of 0.944 with a standard deviation of 0.023. These results highlight the effectiveness of LDA and Isolation Forest in overcoming the challenges posed by low-cost gas sensors, thus enhancing the accuracy and robustness of perfume classification.

Access to Document

Other files and links

• Link to publication in Scopus

Fingerprint