Anomaly Detection On Propulsive Systems By Global Approach Using Autoencoders

Abstract With the increased complexity of space launchers’ missions, propulsive systems’ engines are endowed with evolved control systems for performance optimization and health monitoring. These systems need to operate despite the loss of one or several sensors involved in the control loop. As an alternative to sensor redundancy, validation algorithms based on autoencoders turn out to be efficient to detect drift or failure of sensors and propose reasonable substitution measures. This paper presents the application of the method to such an engine currently in development. It details the autoencoder’s embedding into an overall model providing both detection and reconstruction functions, trained and validated up to high performance directly, without the usual human tuning of thresholds. Performance levels achieved on engine simulation data are displayed.