SMOTE-DNN algorithm for accurate recognition and classification of VAHs with minimal dataset requirements

Volatile aromatic hydrocarbons (VAHs) constitute a major class of gaseous pollutants. Recently, sensor array-based electronic nose (e-nose) technologies have emerged as promising tools for real-time monitoring, facilitating the evaluation of their environmental and physiological impacts. However, compared with other gaseous pollutants, VAHs possess markedly lower chemical reactivity, necessitating elevated operating temperatures and extended sensing durations. These requirements hinder the collection of large-scale datasets, thereby limiting the effective training of artificial intelligence algorithms and ultimately reducing both the accuracy and efficiency of e-nose systems in VAH detection. In this study, we propose a novel Synthetic Minority Over-sampling Technique (SMOTE)-Deep Neural Network (DNN) algorithm for the substantial volume of requisite training data and the associated high costs of data acquisition. Our findings reveal a substantial reduction in the training dataset size, effectively mitigating data collection expenses while achieving comparable performance using only 1% of the original data volume. Notably, this method achieved a classification accuracy of 92.5301% in distinguishing aromatic hydrocarbons from non-aromatic hydrocarbons, and an even higher accuracy of 96.6667% in differentiating among the aromatic species themselves. This study provides a resource-efficient methodology to enhance the precision gas classification tailored to scenarios characterized by limited sample sizes, thereby offering a pragmatic solution to the challenges inherent in real-time mixed gas analysis.