Classification of gases with deep network-based features and regression analysis of concentration values

Electronic nose (e-nose) is a device based on electronic sensing technology that determines, identifies and classifies various properties by making precise measurements on volatile components emitted from biological materials with a chemical sensor array. These systems, which work by imitating the human olfactory mechanism, have been widely used in different fields in recent years. One of the important issues in e-nose is the estimation of different concentration values of different gases. Accurate estimation of gas concentrations plays a crucial role in sensitive issues such as disease detection. This study was carried out to increase the classification and regression success of the concentration levels and values of ethanol, methane, ethylene and carbon monoxide gases detected by 4 metal oxide gas sensors using deep learning networks. Different methods was used to compare the classification successes of gases according to their concentration levels and to compare the estimation successes of concentration values in a certain range. In order to perform these operations, preprocessing and different feature extraction steps were applied to the gas data. The focus of this study is to increase the success of the classification and regression by developing new methods for feature extraction. For this purpose, two new methodologies have been proposed. In one of the developed methods, feature extraction was made from the fully connected layer of Long Short-Term Memory (LSTM) networks, which is one of the deep learning networks, and the extracted features were used in the classification and regression. In the other proposed methodology, the data obtained from the fully connected layer of the LSTM network was applied as an input to the Temporal Convolutional Network (TCN), and the features obtained from the fully connected layer of this TCN were used in classification and regression studies (LSTMFCL-TCNFCL). When the results of both proposed methods are compared with the traditional methods, it is seen that there is an improvement in both classification and estimation results, the success is increased and the mean squared errors are significantly reduced. In addition, it is seen that there is a remarkable increase in the results when signal correction is applied in the preprocessing stage. Both the application of signal correction and the developed methods have provided a bilateral improvement in the results. While the Support Vector Machine algorithm gave the highest accuracy rate with 94,7% in the classification, the lowest mean square errors in regression were obtained with the Gaussian Process Regression.