Quantification of Individual Gas Components for Gas Sensor Array using Neural Networks with Levenberg–Marquardt (LM) and Bayesian Regularization (BR) Technique

Treating the signals from various gas sensors to increase their selectivity and sensitivity is promising. Feed-forward neural networks (FFNN) were used for pattern recognition to discriminate gas species and concentrations. This paper uses a publicly available gas mixture dataset of 16 gas sensor arrays. The dataset contains a mixture of Ethylene & CO and Methane & Ethylene in the air. The FFNN used a variation of Levenberg-Marquardt (LM) and Bayesian Regularization (BR) to train neural networks for better quantifying the individual gas. Based on statistical and computational complexity results, feed-forward neural networks (FFNNs) are the best classifier among the considered training model. Using the FFNN with the LM and BR models, the ppm concentrations of the identified gases were estimated. Both training models are perfect for mixed gas sensor data; Bayesian regularization provides results more accurately than the LR but took a very high run time, infact 3 times the LM model.