Regularized Lightweight One-Dimensional Convolutional Neural Networks for Efficient and Accurate Pollutant Classification

Accurate and efficient classification of air pollutants is essential for electronic nose (e-nose) systems, particularly those deployed on resource-constrained portable devices. Although one-dimensional convolutional neural networks (1-D CNNs) have shown strong potential for gas classification, many existing models are computationally complex and have not been systematically evaluated across diverse e-nose datasets. This study proposes a regularized lightweight one-dimensional convolutional neural network (Reg-LwCNN) for e-nose systems in classifying air pollutants. The proposed model integrates three regularization techniques, i.e., dropout, batch normalization, and weight decay, into a shallow architecture comprising a single convolutional layer, a pooling layer, and a dense layer. To assess its performance in terms of classification accuracy, processing time, and model complexity, Reg-LwCNN was compared with five state-of-theart 1-D CNN models across two public e-nose datasets. The results show that the proposed Reg-LwCNN model achieved the highest classification accuracy (up to 0.986) and the shortest processing time, while maintaining low model complexity across both datasets. Compared to the baseline LwCNN, the proposed model improved classification accuracy by up to 3.03 %, and reduced training time and inference time by up to 32.76 % and 22.27 %, respectively. These findings highlight its potential as a classification algorithm for real-time, edge-based deployment in portable e-nose systems, supporting green information and communication technology (ICT) applications in intelligent environmental monitoring.