A Spatiotemporal Correlation-Based Data Recovery Method for Improving the Performance of Electronic Noses

Issues of sensor failure, data transmission interruption, and even power interruption can result in randomly piecewise missing of detection data of online instrumentation and measurement systems, which will lead to large detection errors and even wrong decisions in real-world scenarios. It has been posed as a critical problem for electronic noses, which have been widely studied as a gas/odor measurement system, as the accuracy of identification and quantification of target gases or odors is greatly dependent on the data acquisition from the sensor array. To address this issue, a novel electronic nose data recovery method based on spatiotemporal correlation is proposed and studied in this work. Specifically, a deep gated recurrent unit (GRU)-based time-series model is devised to explore the potential time-series relationships between the missing data and the historical data. In addition, an extreme learning machine (ELM)-driven spatial correlation model is established to harness the intersensor correlations within the array. Finally, a weighted fusion of temporal and spatial correlation predictions is performed to obtain more accurate recovery data and lower gas detection errors. Experimental validation was conducted using a laboratory-built electronic nose system, which included the detection of peracetic acid concentrations at 0.5% and 0.4%. Experimental results demonstrate a significant improvement in the accuracy of gas concentration quantification when utilizing the proposed data recovery method. Compared with results obtained from random piecewise loss data, both the RMSE and MAE are reduced by at least 50%. It is proved that the proposed approach is capable of effectively recovering data missing occurring simultaneously in the array as well as continuous data missing across multiple time points.