CO Concentration prediction in E-nose based on MHA-MSCINet

Abstract

The prediction of gas concentration plays a key role in human life and health, among which CO is a common toxic gas in industry. In order to protect people's health, the prediction of CO concentration has a worthwhile attention. Electronic nose (E-nose) has performed well in gas concentration prediction in recent years. Among them, the gas concentration prediction performance of E-nose mainly depends on the goodness of the prediction model. Deep learning algorithms can utilize their multilayer networks to extract features from raw data, however, the current application of deep learning algorithms for gas concentration prediction of E-nose is still insufficient, and the prediction results using traditional neural networks often fail to be very fine. Based on this, this study proposes a mish-sample convolution and interaction network based on a multi-head attention mechanism(MHA-MSCINet) for multivariate time series prediction. Our model develops a new module and combines the improved SCINet with the multi-head attention mechanism. Meanwhile, in order to make our model interpretable, we used the SHAP value analysis method. Finally, experiments verify that the model outperforms models such as LSTM, TCN, transformer and SCINet.