Machine Learning Driven Optimization of Gas Sensor Arrays for the Detection of Toxic Industrial Chemicals (TICs)

The detection of toxic industrial chemicals (TICs) such as hydrogen cyanide (HCN), acetonitrile (ACN), methanol (MeOH), and tetrahydrofuran (THF) is essential for ensuring safety in industrial and environmental contexts. However, gas sensors often suffer from cross-sensitivity, signal drift, and inconsistent performance under varying conditions. This thesis presents a machine learning–based framework that enhances TIC detection by leveraging multivariate sensor responses across different temperatures. Sensor data from 160 sensors were collected under exposures to four TICs at three temperature levels. A power-law curve fitting method was used to filter the sensors based on their R² performance, followed by sequential filtering to identify sensors consistently responsive across temperatures and gases. The final reduced dataset was used to train artificial neural networks (ANNs) for gas classification and concentration prediction. Results demonstrate that the ANN model effectively identifies and quantifies TICs using selected high-performance sensors, offering a robust and scalable approach for multicomponent gas sensing in complex environments.