A multi-sensor electronic nose with distribution-aware feature transformation for volatile signatures pattern recognition

This study presents a multi-sensor fusion system that uses feature transformation for volatile signature pattern recognition (VSPR) in ambient-sensing environments. The system enables reliable identification of volatile sources using metal-oxide-based gas sensors and machine learning (ML). An array of eight gas sensors was used to acquire volatile signatures from two four-class datasets representing polycyclic aromatic hydrocarbon (PAH) - related sources and chemical disinfectants. We evaluated five feature-transformation techniques—MinMax Scaler, Standard Scaler, Robust Scaler, Power Transformer, and Quantile Transformer—using five ML classifiers: k-nearest neighbors (KNN), support vector machine (SVM), multilayer perceptron (MLP), logistic regression (LR), and Naive Bayes (NB). t-distributed stochastic neighbor embedding (t-SNE) was used to visualize class separation in two- and three-dimensional space. The five classifiers were evaluated using five-fold cross-validation and standard performance metrics, including accuracy, precision, recall, and F1-score. KNN achieved the best overall performance, with accuracy and F1-scores of approximately 97%–98%. Power transformers and quantile transformers generally improve classification performance. Receiver operating characteristic (ROC) analysis yielded an average area under the curve (AUC) of 0.9925 for Dataset-I and approximately 0.99 for Dataset-II, indicating strong class separability. The Wilcoxon signed-rank test showed significant differences between KNN and the other classifiers ( p < 0.05 ), with large Cliff’s Delta effect sizes ( δ = 0.65 - 0.95). The results highlight the importance of suitable feature transformations in improving volatile-signature classification and demonstrate the potential of multi-sensor fusion for intelligent ambient-sensing applications.