Application of Machine Learning to In2O3-Based Semiconducting Oxide Gas Sensors for High-Performance Gas Discrimination Against Ambient Humidity and Temperature Variations

Abstract

Since the advent of the 4 th industrial revolution characteristic of smart living standards, physical and/or chemical sensors have been gaining their academic/industrial interests in association with cloud-based data management, artificial intelligence and big data thanks to ever-increasing computing power and communication technology. In particular, machine learning-operated sensor networks are advancing to offer predictive, prescriptive, and even deductive analytics, overcoming basic descriptive functions. Regardless of the type of sensor, i.e., physical or chemical, homogeneously and/or heterogeneously configured sensor arrays can provide physical status and chemical information that have been impossible to achieve using single-mode sensors alone. This teaming of technology has opened up unprecedented applications that may be possible through sensor network implementation. Electronic nose with semiconducting gas sensors array can be regarded as a promising platform to find new functionality in the recognition of smells and odors through machine learning. Oxide semiconductor gas sensors with high sensitivity, simple structure, rapid response speed, excellent reversibility and facile integration have been widely employed to detect harmful, explosive, and toxic gases but the simple gas sensing mechanism involving charge transfer between the gas and oxide surfaces often leads to a lack of gas selectivity, hampering gas recognition. The machine learning ecosystem is capable of solving the pre-existing drawbacks encountered in chemical sensor domains. However, the recognition of gases under variations in ambient humidity and temperature has barely been investigated, and most studies have focused on the compensation of sensor signals using humidity and temperature sensor. Gas recognition under various humidity conditions by machine learning without the assistance of humidity sensors has never been achieved. Five In 2 O 3 -based semiconducting metal oxide (SMO) gas sensors were combined in the form of sensor arrays with machine learning methodologies with the aim to detecting and discriminating indoor volatile organic compounds (VOCs) such as benzene, xylene, toluene, formaldehyde, and ethanol against humidity and/or temperature variations. The SMO gas sensor performance was evaluated using principal component analysis (PCA) and neural network-based classification in terms of the gas sensor data type/amount, neural network algorithms, sensor combinations, and environmental factors. The PCA analyses revealed the limitations on the discrimination of VOCs under temperature- and/or humidity-interfered gas sensing environments. Gas detection/discrimination could be improved significantly by using neural network-based algorithms, i.e., artificial neural networks (ANNs), deep neural networks (DNNs), and 1-dimensional convolutional neural networks (1D CNNs). The neural network algorithm prediction based on the entire gas sensing/purge transient data outperforms deep learning-assisted predictions based on partial gas sensing transients. Compared to 1D CNNs, DNNs are more appropriate in terms of training/validation/test datasets. The effects due to humidity variation are proven to more significant than those due to temperature fluctuation. A 2-sensor mode combination can be exploited to replace the 5-sensor operation in ML-based applications without significantly losing the prediction accuracy. The indoor VOC pollutants can be successfully discriminated even under the variation of ambient humidity and/or temperature by ML-based approaches. The 2-sensor mode concept can be exploited as a super-sensor strategy applicable to the practical domestic and industrial sites. The implications and limitations of machine learning are discussed in terms of mutually-interacting SMO gas sensors.