9:15 AM - 9:30 AM
▼ [24a-E105-2] Identification of Gas Mixture by A Thin Lines SnO2 Gas Sensor Using Convolution Neural Network with Fragments Blend and Unified-Weight
Keywords:semiconductor oxide, gas sensor, CNN
The elevated acetone level in human’s breath is a biomarker for indicating diabetes mellitus. Smart gas sensors associated with machine learning approaches are promising to recognize target gas from a complicated background in respiration. However, it could be hard to build a database with all of gas types and concentrations. Thus, gas mixture recognition using the minimal dataset has been evaluated in this work, aiming at relieving the workload of data collection.
1 μm-width thin lines SnO2 gas sensor was applied to record the resistance fluctuations under acetone and acetone/ethanol mixtures with a sampling rate of 10 kHz. 56700 seconds resistance data containing 12 types of gas components were obtained for the data processing, and the data of individual acetone were used to create the minimal training datasets. The convolutional neural networks with fragments blend and unified-weight (CNN-FBUW) were proposed and compared with the traditional CNN. To reduce the risk of overfitting, the selected data was divided into fragments to increase data amount and we found the optimal data length (interval) is 2000. Furthermore, the unified weight of each fragment’s results was introduced for preventing the interference from environmental noise. The average accuracy of acetone recognition in mixture reached the highest when using 1D CNN-FBUW.
In this work, the recognition of gas mixture was demonstrated by using one thin lines SnO2 gas sensor and deep learning approaches. It is possible to recognize acetone concentrations from the acetone/0.5 ppm ethanol and acetone/1 ppm ethanol mixtures with the model only trained with individual acetone data.
1 μm-width thin lines SnO2 gas sensor was applied to record the resistance fluctuations under acetone and acetone/ethanol mixtures with a sampling rate of 10 kHz. 56700 seconds resistance data containing 12 types of gas components were obtained for the data processing, and the data of individual acetone were used to create the minimal training datasets. The convolutional neural networks with fragments blend and unified-weight (CNN-FBUW) were proposed and compared with the traditional CNN. To reduce the risk of overfitting, the selected data was divided into fragments to increase data amount and we found the optimal data length (interval) is 2000. Furthermore, the unified weight of each fragment’s results was introduced for preventing the interference from environmental noise. The average accuracy of acetone recognition in mixture reached the highest when using 1D CNN-FBUW.
In this work, the recognition of gas mixture was demonstrated by using one thin lines SnO2 gas sensor and deep learning approaches. It is possible to recognize acetone concentrations from the acetone/0.5 ppm ethanol and acetone/1 ppm ethanol mixtures with the model only trained with individual acetone data.