An Encoder-Decoder Structure With Strong Resistance to Sensor Drift in Machine Olfaction

During the long-term development of gas sensors, it has been challenging to improve the recognition accuracy of the sensors. Since gas sensors are susceptible to aging, chemical poisoning and other factors, the collected data are prone to drift. The paper proposed a Bi-GRU-based Encoder-Decoder framework and introduced a spatio-temporal attention mechanism (SABG for short). The algorithm extracts the feature information of e-nose data through both temporal and spatial dimensions, which can effectively counteract the sensor drift phenomenon with the highest classification accuracy of 91.45%. To demonstrate the reliability of the algorithm, this paper uses a publicly available dataset produced by the University of California, Irvine (UC Irvine or UCI) for experiments. In addition, this paper compares various advanced gas sensor drift compensation algorithms such as PCA, CC-PCA, OSC, SVM-rbf, etc. The experimental results show that the SABG model is able to simulate the dynamic behavior of the input sequence, thus improving the accuracy and generalization ability. It also effectively captures the past and future contextual information to resist the effects caused by sensor drift, which helps the model make more accurate predictions.