Description
A Master of Science thesis in Engineering Systems Management by Lubna Syeda Mahmood entitled, “Assessment and Performance Analysis of Machine Learning Techniques for Gas Sensing E-nose Systems”, submitted in November 2021. Thesis advisor is Dr. Zied Bahroun and thesis co-advisor is Dr. Mehdi Ghommem. Soft copy is available (Thesis, Completion Certificate, Approval Signatures, and AUS Archives Consent Form).
Abstract
The electronic nose, commonly known as the E-nose that combines gas sensor arrays (GSAs) with machine learning, has gained a strong foothold in gas sensing technology. The E-nose, inspired from the human olfactory system, is used for the detection and identification of various volatile organic compounds (VOCs). GSAs produce a unique signal fingerprint for each gas, providing vital information for machine learning algorithms to detect the gas type using classification and estimate its concentration through regression. The inexpensive, portable and non-invasive characteristics of E-noses have rendered them indispensable within the gas-sensing arena. As a result, E-noses are now widely employed for several applications in food industries, disease diagnosis, and environment management. In this thesis, we first review various sensor fabrication technologies and provide a comprehensive literature review of machine learning in gas sensing. Then, we present a detailed assessment of machine learning models employed for classification and regression using the software tool RapidMiner. The models discussed in this thesis include the Artificial Neural Networks, k-Nearest Neighbors, Decision Tree, Random Forests, Support Vector Machine and other ensembling-based models. The models are tested on three different experimental datasets obtained from MoX gas sensors as reported in the literature, followed by their performance analysis. The obtained results are compared against those reported in previously published studies. Classification accuracies reached 99.38% using Random Forests and Support Vector Machine whereas mean absolute percentage errors (MAPEs) were found as low as 5.98%, 8.89%, 6.35% using the k-Nearest Neighbors, Random Forests, and ensemble methods, respectively. Techniques of feature selection and Principle Component Analysis (PCA) retained significant signal characteristics that improved model performances with MAPEs of 8.15% using k-Nearest Neighbors and 4% using Random Forests. The assessment, thus, highlights factors that play a pivotal role in machine learning for gas sensing and sheds light on the predictive capability of different machine learning approaches applied on experimental GSA datasets.