Batch Normalization and Dropout Regularization in Developing Lightweight Deep Learning Models for Electronic Nose Gas Classification

Abstract

Developing lightweight deep learning classification algorithms for accurate and efficient air pollutant classification is crucial for electronic nose (e-nose) systems. Although batch normalization and dropout are widely used in deep learning, their impact on classification accuracy and model efficiency in e-nose-based lightweight deep learning models remains underexplored. This study evaluates the effects of these regularization techniques in lightweight multilayer perceptron (MLP) and one-dimensional convolutional neural network (1DCNN) models using two public e-nose datasets. Each model was implemented with and without regularization to assess classification accuracy and model efficiency. Results show that while the baseline 1DCNN achieved better accuracy and efficiency than the baseline MLP across both datasets, statistical analysis found no significant difference between the two models. However, the findings also show that adding batch normalization or dropout to the MLP significantly degraded its performance relative to the 1DCNN, whereas adding dropout to the 1DCNN yielded significant improvements over MLP-based models. Thus, this study provides a systematic evaluation of batch normalization and dropout regularization in developing lightweight deep learning models for smart e-nose gas classification systems.