A Resilient to Faults Auto-Encoder Enabled Kalman based Multi-Sensorial Fusion

This article presents a novel Auto-encoder-enabled fault resilient multi-sensor fusion architecture while incorporating an extended Kalman filter framework. The auto-encoder facilitate reconstruction of the faulty measurements from multiple onboard sensors, while the centralized extended Kalman filter enables an accurate fusion architecture. Moreover, the process is capable of successfully eliminating the additive noise appearing from the raw sensor data. The proposed method provides a robust reconstruction mechanism in the presence of time-dependent anomalies and faulty sensor measurement. The efficacy of the proposed scheme is extensively evaluated in the context of pose estimation for a micro aerial vehicle equipped with multiple onboard sensors. In addition, the evaluation process incorporates various realistic failure scenarios with artificially introduced inaccurate measurements. The superiority of the proposed Auto-encoder enabled centralized Kalman filter (AEKF) fusion is demonstrated through an extensive comparison with a recently developed Fault Resilient Optimal Information Filter (FROIF) method.