Multisensor Detection Design via a Weighted Scheme With AUC and Information Theory

In this article, a multisensor detection system is constructed for Herbal Medicine odor detection and recognition based on a weighted scheme with area under curve (AUC) and information theory. In the feature extraction stage, the multifeature combination that is more conducive to classification is selected by comparison to lay a good foundation for subsequent sensor array optimization. In the array optimization stage, the performance of the array optimization algorithm on unbalanced datasets is further improved by developing the joint mutual information (JMI) and AUC measure (AJMI) based on the consideration of correlation and redundancy among features. Experimental results demonstrate the superiority of the proposed design in comparison to other methods from the literature. In particular, it is verified that the proposed AJMI algorithm can improve recognition accuracy by 4% and F1 score by more than 4% while reducing the number of sensors from 37 to 23 compared with the multiple-feature extraction design without considering array optimization. It is worth noting that the proposed design achieves better recognition results than other algorithms when using only 14 features from ten sensors.