Identification of binary gases’ mixtures from time-series resistance fluctuations: A sensitivity-controllable SnO2 gas sensor-based approach using 1D-CNN

The elevated acetone levels in human breath could be a promising biomarker to predict diabetes non-invasively, while requiring a low concentration of target gas detection among complex respiration background gases. This research presents highly selective detection of acetone and its binary mixture in the pure air by a single SnO2-based metal oxide gas sensor. The sensor, whose sensing area was optimized by the nanometric thin lines structure to increase its surface-to-volume ratio, was combined with a 1D-convolution neural network (1D-CNN) and optimized with fragments blend and unified weights (CNN-FBUW) to identify acetone concentrations from time-series resistance noise. To explain the recognition mechanism, we have transformed the input data and trained those datasets with the same CNN architecture. Amplitude and high-frequency information of resistance fluctuation were proven essential for gas recognition. A higher sensitive device amplifies the amplitude information, benefiting deep learning algorithms to explore adsorption and desorption processes on the SnO2 layer. Acetone/ethanol binary gases’ mixtures were possibly identified with the model only trained with individual acetone gas data, which is considered a novel approach to relieving the data collection workload. [Display omitted] •An intelligent system was developed to recognize gas mixtures based on a thin lines SnO2 gas sensor and 1D-CNN.•The sensitivity-controllable sensors were used to reveal the relation between the performances of sensors and recognition.•Deep learning models were developed to recognize gas mixtures from resistance fluctuations with the minimal dataset.•The proposed sensing system can be applied to recognize other disease-specific volatile organic compounds.