Statistical shape analysis pre-processing of temperature modulated metal oxide gas sensor response for machine learning improved selectivity of gases detection in real atmospheric conditions

Abstract

Development of new signal processing approaches is essential for improvement of the reliability of metal oxide gas sensor performance in real atmospheric conditions. Advantages statistical shape analysis (SSA) method are presented in comparison to previously reported signal pre-processing techniques – principal component analysis (PCA), discrete wavelet transform (DWT), polynomial curve fitting (PCF) – used in combination with machine learning (ML) algorithm for improvement of detection selectivity. An enhanced identification of chemically related gases (methane and propane) at a concentration range of 40−200 ppm under variable real atmospheric conditions has been demonstrated using working temperature modulated metal oxide gas sensors. Laboratory samples of sensors based on nanocrystalline SnO2 modified with Au and Pd were used. The proposed data pre-processing algorithm is less sensitive to sensor response and baseline drift and fluctuations compared to other methods during two months of continuous operation and work with periods of inactivity. The collected dataset and signal processing code are made public. The advantages of SSA signal pre-processing method are also demonstrated with the use of independent publicly available dataset for the task of CO selective quantitative detection in the air with variable humidity in the 2.2−20 ppm concentrations range.