A Rocket Payload Demonstrator for Real-Time Fault Detection of Pressure Sensors Based on Redundancy and Machine Learning

We report on a novel sensors fault detection electronic system combining physical redundancy with machine learning. An unsupervised one-class support vector machine classifier analyzes the time evolution of the signals of three solid-state pressure sensors operating in parallel and provides a healthiness score to a voting system able to estimate the correct value of the measurand even in the presence of a majority of faulty sensors. The system is built into a single-unit Cubesat format to be launched as the technical payload in an experimental competition rocket developed by students in which the barometric pressure is used to determine the rocket altitude, in particular when the target apogee is reached. The system design and simulations, showing 70% accuracy in fault detection and resilience to multiple faults, are here presented.