Missing Sensor Data Restoration: Computationally Intelligent Discovery of Reading Dependencies

University Of Washington, Seattle WA

Investigators

Abstract

0114483 El-Sharkawi This project will explore the possibility of using new techniques from computational intelligence in order to address the task of imputing missing values in data collected by complex arrays of sensors. Sensors in a restricted environment can have readings relating to each other in such a matter that missing data from a set of failed sensors can be restored from the measurement of those remaining. When the reconstructed sensor readings are supplied as input to a process, the operation of the system is therefore still possible. Overall performance will degrade gracefully, if at all. Discovery of data constraints can be achieved by training an autoencoder using sensor data. The trained encoder, having empirically discovered interrelations among data, can then be used to restore lost readings from failed sensors. This is achieved through either application of an alternating projection onto convex set (POCS) algorithm or a more conventional search. Successful development of missing sensor data (MISED) restoration will have significant impact on a number of current technologies. The project will focus on applications to currently open and important problem in energy and avionics. The project has strong support from both industries.