<![CDATA[Objective coffee aroma evaluation remains challenging outside controlled laboratory settings, and most electronic nose studies neglect embedded deployment constraints. This work proposes an edge-aware coffee aroma classification framework that integrates multi-representation feature extraction with LightGBM and evaluates both predictive performance and computational efficiency. A six-sensor metal-oxide semiconductor (MOS) e-nose was developed, producing a balanced dataset of 1,080 trials from 12 aroma classes. Five feature representations were investigated, including baseline signals, autoencoder embeddings, and convolutional features derived from pseudo-image transformation. Experiments on an NVIDIA Jetson Nano using stratified five-fold cross-validation showed that residual-based representations significantly improved performance. The lightweight residual network achieved an accuracy of 0.9972 with low training time and memory usage. Pareto analysis confirms that optimal performance is achieved by balancing accuracy and resource constraints, thereby enabling reliable deployment in edge and IoT environments.]]>
Copyrights © 2026
