Design and Implementation of a ResNet-LSTM-Ghost Architecture for Gas Concentration Estimation of Electronic Noses

Gas quantification has been proved an effective function of electronic noses, and how to improve the accuracy of concentration estimation has been a critical issue now. Aiming for this, a high-precision gas concentration estimation model is proposed based on a deep-learning framework in this work. A 2-D texture image feature containing the spatial-temporal information of the gas sensor array is first obtained and a ResNet-LSTM-Ghost architecture is correspondingly designed and implemented for the purpose of gas concentration estimation. The spatial-temporal features are learned by the ResNet, the long-term dependency between time dependent sensor responses is modeled by a long short-term memory (LSTM) network, and the Ghost module is embedded to extract deeper spatial-temporal information in the feature matrix. Different concentrations of four kinds of disinfectants have been measured by an array of ten sensors. Experimental results show that gas concentration quantification coefficient of determination (
${R} ^{{2}}$
) of the proposed method is improved at least 0.0409, and the relative error (RE) of leave-one-group-out experiment is reduced by a minimum of 1.64%, compared with five classical machine-learning algorithms.