A 2D-CNN Architecture for Improving the Classification Accuracy of an Electronic Nose with Different Sensor Positions

Responses of metal oxide (MOX) gas sensors are affected by various factors, such as their locations relative to the gas path. For achieving good classification accuracy by an electronic nose (e-nose) which involves similar sensors in different places, it is significant to extract informative features according to the spatial information of signals. A popular e-nose dataset, including the responses of 72 MOX sensors of eight types and each one at nine different positions, was used to investigate the importance of considering the location of sensors in classification algorithms. A method is proposed using a simple two-dimensional convolutional neural network, 2D-CNN, and based on spatial and temporal information of sensors. Then, a 1D-CNN is designed without the location information of sensors and with almost the same number of parameters. It is shown that the 2D-CNN scheme results in 97.8% detection accuracy, which is 7.5% upper than the accuracy of 1D-CNN. In conclusion, better feature extraction and more accurate classification can be obtainable through spatial and temporal information in 2D-CNN.