An Algorithmic Approach to the Optimal Extraction of Signals from Intelligent Sensors

This paper describes the development of an intelligent sensor architecture, where signal conditioning is performed onboard the sensor itself, in software. Our proposed architecture uses data-based models of the sensor for signal conditioning and fault detection, so that the sensor is robust to degradation and its processed output includes an estimate of uncertainty with each measurement value for higher level sensor management processes such as data fusion. We use a data-based kernel representation for the signal conditioning system, which avoids deriving physical models of the sensor from first principles. A sparse realisation of the kernel model provides fast predictions and opportunities for efficient updating of the sensor model to enable reconfiguration of the sensor model based on incoming data. We show that these techniques have the ability to detect degradation in a MEMS sensor, using elevated temperatures in laboratory conditions