Application of Unsupervised/Supervised Machine Learning to Harmful Gas Detection/Classification Based on Mutually-Interacting Semiconductor Metal Oxide Gas Sensors

Abstract

Chemical and/or physical sensors are employed at domestic and industrial sites with the aim to perform a variety of functions, such as prediction, forecasting, prognostics, remaining useful life estimation, anomaly detection, and trend analysis. The sensing information contain the analog signals originating from gas-generated signals, pressure, shear, strain, torsion, temperature, humidity, illumination, electromagnetic radiation, sound, and vibration. Chemical gas sensors can be categorized mainly as semiconductor metal oxide (SMO), electrochemical, or photo-ionization detection (PID) sensors. SMO gas sensors are based on adsorption and desorption of gas molecules onto/from semiconductor metal oxides and however, suffer from the low selectivity to diverse gas species, i.e., the mutually interacting features.Here, we report a high-performance machine learning strategy for gas detection/discrimination against harmful gas combination and adopt the synergistic integration of supervised and unsupervised learning by exploiting a SMO gas sensor array. A 4 SMO sensor array was constructed for detecting carbon monoxide and ethyl alcohol (C2H5OH) mixtures using 15 different gas combinations. Gas detection/discrimination performance was probed in terms of different numbers of gas sensor integrated as an array and machine learning algorithms. Unsupervised K-Means clustering was successfully applied to the rational identification of the similarity features of targeted gases among 4 different groups, which are listed as matrix gas, two single-component gases, and one two-gas by employing only unlabeled voltage-based gas sensing information. Detailed classification was performed through a multitude of supervised algorithms, i.e., 2-layer artificial neural networks (ANNs), 4-layer deep neural networks (DNNs), 1-dimensional convolutional neural networks (1D CNNs), and 2-dimensional convolutional neural networks (2D CNNs). The numerical-based DNNs and image-based CNNs are evaluated to be excellent approaches for gas detection/classification, as corroborated by the highest accuracy and lowest loss parameters.Through the systematic investigation on the influence of the number of sensors on the arrayed gas sensor system, the applicability of machine learning methodology to an arrayed SMO gas sensor system is justified by the four unique features, i.e., i) a data augmentation methodology, ii) machine learning approach of combining K-means clustering and neural networks, and iii) a systematic approach to optimized sensor combinations, potentially recommending the minimized sensor systems based on chemical gas sensors against harmful gas species. Even two SMO sensor combinations are shown to be highly effective in performing gas discrimination against harmful gas species assisted through either numeric-based DNNs or image-based 1D CNNs, overcoming the simple clustering proposed through the unsupervised K-means clustering.