Minimizing sensors for system monitoring - a case study with EEG signals

For monitoring any system, be it a chemical plant, a nuclear power station, or a human heart or brain, we need to attach sensors and analyze the multivariate time-series data collected by those sensors. If the system has two states, say state 1 (good) and state 2 (bad), we need to infer which state the system is, by classifying the collected time-series signals. To make this work efficiently, it is important to search the least number of probes that would give best classification result. It is a multi-objective optimization problem. The proposed approach works in two steps. We start with a large number of probes. As the first step, we cluster the time-series signals. and choose a representative one from each cluster. Next, we run pareto GA to select the smallest set of probes (from cluster representatives), that would give the highest classification result. Depending on the nature of the signals, and the target application, appropriate signal-features, clustering and classification algorithms will be different, but the basic principle is applicable to any system. In this paper, we tested the effectiveness of our algorithm with EEG signals, to detect the presence or absence of ERP 300. Improved results with less number of probes compared with previous works validated the approach.