Using transfer-learning-based algorithms as data reduction strategies for volatile organic compounds classification using plasma spectroscopy

Abstract

This work presents the use of transfer-learning-based algorithms as data reduction strategies for the classification of volatile organic compounds (VOCs) using the optical emission spectroscopy of plasmas. The plasma used is generated with a home-made microplasma generation device (MGD) ignited in the mixtures of Ar and VOCs. The spectra are acquired from ten MGDs. The VOCs tested are methanol, ethanol, and isopropanol. VOCs are classified using a convolutional neural network. In addition, gradient-weighted class activation mapping is used as the explainable artificial intelligent technique. It ensures the model classification is based upon rational plasma physics by considering appropriate wavelengths. The VOC concentrations are then quantified using linear regression and an artificial neural network (ANN). The transfer learning-based algorithms tested are parameter transfer, REPTILE, and self-training. Spectral data from ten MGDs are grouped into source and target datasets. Ten MGDs are tested individually using a model that was trained on the other nine MGDs. The three MGDs with the lowest accuracy are chosen as the target dataset, while the other seven MGDs make up the source dataset. The original target dataset has 22 500 spectra and is further reduced to 12 600, 9000, 1800, 225, and 22 spectra to test the behavior of each algorithm. With 225 spectra used for training, the model trained with the random initial model shows an accuracy of 0.82. The models trained with parameter transfer and REPTILE have accuracies of 0.98 and 0.95, respectively. Finally, an ANN model is used to quantify the VOC concentration with an R 2 value of 0.9996. The results demonstrate the potential using transfer-learning-based algorithms as the data reduction strategies for classification of spectroscopic data.