Neural Network-Based Prediction of Response Signal of Metal Oxide Semiconductor Gas Sensors

This study presents a novel method for predicting the response signal and recovery time of metal oxide semiconductor (MOS) gas sensors at different gas concentrations just based upon their initial power-on behaviors in air. First, we measured the resistance changing behavior of the MOS gas sensors during the power-on period in pure air (power-on data). Second, their response behaviors, including response signal as well as recovery time, in the target hydrogen gas of varying concentrations (from 20 to 1000 ppm) were collected (signal data). The initial power-on data and the signal data were found to be closely related based on a neural network model, therefore one can use just the power-on data to predict the gas sensor’s signal in the target gas at different concentrations. Thus, the tedious calibration work for these MOS gas sensors in real target gas can be dispensable. Two types of neural networks were used for the model: Artificial Neural Network (ANN) and Convolutional Neural Network (CNN). Experimental results indicate that the CNN outperforms the ANN in both response signal and recovery time predictions, with an average voltage prediction error of 0.166 V and an average recovery time prediction error of 4.746 s. Instead of using measurements in actual gases, this study offers a practical way to obtain the signal data (i.e., response signal and recovery time) of MOS gas sensors.