Neural Network Model for Gas Classification of Semiconductor-Based Heterogeneous Gas Sensors Arrays

The neural network is proposed for the classification of six different gases (ammonia, acetaldehyde, acetone, ethylene, ethanol and toluene). The proposed network is trained, validated and tested on the 13,910 dataset points having 129 feature sets of parameters. These parameters are generated from commercial chemical sensors at different concentration levels. Initially, the database is analysed to compute the correlation between data points. The proposed neural network has three layers with different activation functions and different number of nodes. The input layer has 32 nodes, the hidden layer has 16 nodes and the output layer has 6 nodes. The network is trained in a batch size of 8 data points for 70 epochs. The model loss and accuracy for the training and testing phases are plotted. Finally, the confusion matrix of six gases is presented in the paper. The 746, 808, 447, 568, 872 and 524 data points of gas ammonia, acetaldehyde, acetone, ethylene, ethanol and toluene respectively are classified correctly. The proposed model can be enhanced and trained for other similar gases using a transfer learning approach.