A Censoring Scheme for Multiclassification in Wireless Sensor Networks

Censoring has been widely applied in wireless sensor networks (WSNs) as an effective method to achieve a balance between energy consumption and the quality of the observed signals. However, most recent studies focus on the censoring schemes applied in binary hypothesis problems (i.e., binary classification and detection problems). To expand the application of censoring in WSNs, we propose a censoring scheme for multiclassification problems in this article. Sensors in this scheme only transmit observations deemed informative enough for classification, where the decision region of whether to transmit is derived based on log likelihood ratios (LLRs). By analyzing the relationship between the communication rate of the WSN and the censoring threshold, we design an adaptive strategy in the censoring scheme so that the censoring threshold can be adjusted according to the communication rate. We further derive the theoretical lower bound of the classification accuracy, which is formulated via the Chernoff distance among different signals. The performance superiority of the censored signals compared with the original ones without censoring is revealed in the form of the theoretical lower bound, verified by experimental results on WSN applications where our proposed censoring scheme allows significant communication saving without the sacrifice of performance.