Do I look like my neighbors? Methods to quantify sensor data similarity

Type: Semester thesis
Student: Sarah Neff
Advisor: Claudia Villalonga, Dr. Daniel Roggen
Project: OPPORTUNITY, SENSEI

With the increasing of the number of sensors, sensor replacement becomes a key issue to ensure fault tolerance of the system. Thus, for a new sensor joining a network, it is important to know if it provides a similar data than one of its neighbors, so that it could replace them in case of fault.
In this thesis, we propose a method to detect sensor data similarity: the sensors are classied in two classes, either close or far. Thus, two close sensors provide similar data and could replace each other. The method is optimized to take few computational time, by using low complexity features (mean, variance, mean-crossing rate and RMS) and an Euclidean distance. It achieves low latency and high accuracy in proximity detection on the BA Fitness data set. It is furthermore validated on the Opportunity data set, to detect human object interaction, and on a newly recorded data set where two people interact with the same object. In both cases, the method proves good results, achieving accuracies above 80% for a window size of 6s.