Reliable Sensor Intelligence in Resource Constrained and Unreliable Environment

Abstract

The objective of this research is to design a sensor intelligence that is reliable in a resource constrained, unreliable environment. There are various sources of variations and uncertainty involved in intelligent sensor system, so it is critical to build reliable sensor intelligence. Many prior works seek to design reliable sensor intelligence by developing robust and reliable task. This thesis suggests that along with improving task itself, task reliability quantification based early warning can further improve sensor intelligence. DNN based early warning generator quantifies task reliability based on spatiotemporal characteristics of input, and the early warning controls sensor parameters and avoids system failure. This thesis presents an early warning generator that predicts task failure due to sensor hardware induced input corruption and controls the sensor operation. Moreover, lightweight uncertainty estimator is presented to take account of DNN model uncertainty in task reliability quantification without prohibitive computation from stochastic DNN. Cross-layer uncertainty estimation is also discussed to consider the effect of PIM variations.