A novel pathway to construct gas concentration prediction model in real-world applications: Data augmentation; fast prediction; and interpolation and extrapolation

Abstract

With a growth in gas sensor applications in various fields, high-performance gas sensors are required. In particular, an accurate prediction of gas concentration is needed for monitoring an indoor/outdoor air quality, inspection of food freshness, and medical diagnosis. Developing a gas concentration prediction model for real-world application requires three critical features: sufficient training data, rapid prediction, and accurate prediction of gas concentrations. Here, we propose techniques to fulfill these requirements. For fast prediction, we employ the initial slope of response for prediction instead of the widely used steady-state response. Using the 1/f noise measured from the sensor, we augment response data to provide sufficient data to train the prediction models. Finally, we adopt extrapolation and interpolation performances as new metrics to evaluate the performance of gas concentration prediction model. By combining the aforementioned methods, the Linear Regression successfully predicts 6 gas concentrations outside the training data using only 5 gas concentrations as training data.