Ultra-selective tin oxide-based chemiresistive gas sensor employing signal transform and machine learning techniques

Selective detection of gases has been a major concern among metal-oxide based chemiresistive gas sensors due to their intrinsic cross-sensitivity. In this endeavor, we report integration of single metal-oxide based chemiresistive sensor with different soft computing tools to obtain perfect recognition of tested analyte molecules by means of signal processing, feature extraction and machine learning. The fabricated sensor device consists of SnO2 hollow spheres as the sensing material, which was synthesized chemically. A remarkable gas sensing performance has been observed towards every target volatile organic compound (VOC); which exhibits the sensor having cross sensitivity. The transient response curves obtained from the sensor were processed using fast Fourier transform (FFT) and discrete wavelet transform (DWT) to squeeze out distinct characteristic features associated with each tested VOC. The signal transform tools were taken in a comparative fashion to examine their credibility in terms of feature extraction and assistance for pattern recognition. The extracted features were assigned as input information to the machine learning algorithms in a supervised manner to discriminate among the tested VOCs qualitatively. Moreover, a quantitative estimation of concentration for corresponding VOCs was also obtained with acceptable accuracy. The main highlight of the paper is the vigilant and efficient selection of features from the transformed signal which adequately allows the machine learning algorithms to achieve excellent classification (best average accuracy: 96.84%) and quantification. The collective results promote a step towards the realization of an automated and real-time detection.

Automated summary

In this study, a metal-oxide based chemiresistive sensor was integrated with soft computing tools for selective detection of gases. The sensor exhibited remarkable sensing performance towards various volatile organic compounds (VOCs) with the help of signal processing, feature extraction, and machine learning. The extracted features were inputted into machine learning algorithms to qualitatively discriminate among different VOCs and quantitatively estimate their concentrations. The approach achieved excellent classification accuracy (best average accuracy: 96.84%).