Electronic-nose (e-nose) arrays are often over-provisioned, inflating energy use, memory footprint, and latency—constraints that are particularly critical on edge devices. This study formulates sensor selection as a multi-objective problem and introduces a hybrid workflow that couples Kernel Principal Component Analysis (KPCA) with complementary feature selection to obtain compact, accurate models. Using an eight-sensor MOS array, we collected roasted-coffee data spanning 12 aroma classes (90 samples per class). Signals were standardized and transformed using KPCA (RBF kernel). Feature pruning involved combining supervised Recursive Feature Elimination (RFE) with an unsupervised information-content criterion, and the final subset was derived using the Non-dominated Sorting Genetic Algorithm II (NSGA-II), which jointly optimized accuracy, sensor count, runtime, and model complexity. Five repetitions with stratified 5-fold cross-validation showed low variability, with CV-accuracy standard deviations ranging from ≈0.05 to 0.12 across configurations. A Pareto-front analysis revealed a knee near six sensors: KPCA+RFE achieved 99.54% accuracy with shallow trees, while KPCA+RFE+NSGA-II reached 98.61% accuracy with the lowest runtime (0.2126 s) and reduced depth (12.89). Both hybrid approaches significantly outperformed PCA+RFE and random subsets (paired t-tests, p < 0.05). These results indicate that the proposed pipeline delivers edge-oriented efficiency—lower runtime, fewer sensors, and simpler models— without significantly compromising recognition performance. The framework is directly applicable to portable e-noses and can generalize to other VOC tasks; future work includes hardware deployment and multi-classifier validation.
Affiliations come from OpenAlex; journal OAI-PMH metadata carries none. Authors without an OpenAlex ID are matched by name, which can merge or split people.