FFT analysis of temperature modulated semiconductor gas sensor response for the prediction of ammonia concentration under humidity interference

The increasing environmental contamination forces the need to design reliable devices for detecting of the volatile compounds present in the air. For this purpose semiconductor gas sensors, which have been widely used for years, are often utilized. Although they have many advantages such as low price and quite long life time, they still lack of long term stability and selectivity. Namely, environmental conditions have significant effect on the sensing accuracy. That is caused by the fact that sensors also respond to interfering molecules coexisting in ambient gas (e.g. humidity) and their response is highly dependent on the temperature and the gas delivery rate. Among the different strategies used to overcome those shortcomings, the modulation of the sensors' operating temperature has been reported. To perform the interpretation and extraction of useful information from dynamic nonlinear response of temperature modulated sensor, the feature extraction and data processing methods are required. In this article the method of determination the concentration of ammonia in the presence of relative humidity is presented. For this purpose the operating temperature of a single commercial SnO2 gas sensor is modulated using sinusoidal voltage applied to the heater. The measurements are performed for different concentrations of ammonia at specified levels of relative humidity. The validation data set was obtained 100 days after the set used for the calibration data. Several features from the dynamic measurements are extracted. Qualitative and quantitative analysis of the selection of the signal features containing useful and relevant information for prediction of target gas are performed. The assessment of the impact of the input vector length of the Fast Fourier Transform method on the resulting signal features is examined. The selected features are utilized as an input for Partial Least Squared regression. The calibration is performed and the prediction error of the ammonia concentration is calculated based on the validation data.