A fast and robust mixture gases identification and concentration detection algorithm based on attention mechanism equipped recurrent neural network with double loss function

In the daily life, there is a tremendous need of fast monitoring of mixture gases and corresponding concentration estimation. For example, in the petrochemical industry, toxic and flammable gases, such as carbon monoxide, methanol and ethanol always co-exist. Without fast and accurate gas species identification and concentration detection system, great security risks will exist. Here, we propose to use an electronic nose system consisting of attention mechanism equipped recurrent neural network with double loss function (2L-ARNN) algorithm to achieve above tasks simultaneously. Firstly, an end-to-end encoder and decoder frame is applied to offer the flexibility to process variable length of the input. Then a novel gated recurrent unit network facilitates the automatic feature extraction from temporal dynamic behaviour. Based on this, attention mechanism dynamically assigns weight vectors of the gas features, which strengthens the robustness for the concentration fluctuation scenarios. Finally, with double loss function, both targets of mixture gas identification and concentration calculation are achieved using the same network. Consequently, within 5 s, 2L-ARNN algorithm has probably an accuracy of 97.67 % for air, CO, ethylene and methane mixture gases, which is significantly improved compared to classical methods. Simultaneously, the estimated normalized root mean squared errors for CO, ethylene, methane are 6.20 %,7.86 % and 5.71 %, which is 6% smaller than that of convolutional neural network. Therefore, our proposed algorithm features fast processing rate, superior accuracy for both gas identification and concentration estimation with improved anti-interference capability for wide-range of mixture gases conditions are essential for real-time gas monitoring systems.