A Fast and Robust Gas Recognition Algorithm Based on Hybrid Convolutional and Recurrent Neural Network

Fast recognition of flammable and toxic gas species within very short response time is a challenging task for the gas sensing devices adopted in a wide range of applications. The recognition accuracies of the previous implementations are always constrained by the limited feature or dynamic information extracted from the short transient gas response curves. In order to address this issue, in this paper, we propose a novel hybrid approach with both convolutional and recurrent neural networks combined, which is based on the long short-term memory module. Featuring the capability of learning the correlations of time-series data, the proposed deep learning method is well-suited for extracting the valuable transient feature contained in the very beginning of the response curve. As a result, within a response time as short as 0.5 s, the proposed implementation is capable of recognizing the gas species with an accuracy of 84.06%. In addition, the aforesaid accuracy can be further improved by increasing the response time with the step of 0.5 s. According to our extensive experimental results, the recognition accuracy can be elevated up to 98.28% at the response time of 4 s, where it typically needs 40 s for the response curve to achieve saturation. The reported accuracy dramatically outperforms the previous algorithms, including gradient tree boosting (GTB), random forest (RF), support vector machine (SVM), k-nearest neighbor (KNN), and linear discriminant analysis (LDA). Moreover, various standard drift related experiments are conducted, of which the results validate our proposed algorithm's superior robustness for the wide range of real-life applications.