Machine Learning with Echo State Networks for Automated Fault Diagnosis in Small Unmanned Aircraft Systems

Echo State Network (ESN) is one of machine learning methods that can be used to detect anomalies in sensor readings. The method predicts output signals, from which a prediction error can be created. To enable fault-tolerant control, ESN needs to be combined with a robust fault estimation method. Indeed, identifying the source of the faults, whether coming from sensors or actuators, is crucial in safety-critical Unmanned Aircraft Systems (UAS), since it will determine proper control actions when the faults occur. This paper presents a novel method to combine sensor anomaly detection using ESN with actuator fault estimation using adaptive extended Kalman filter (AEKF). Numerical results show the benefit of using the cascaded algorithm in a noisy environment. Furthermore, the presented method is validated using a hexacopter with actuator faults in indoor experiments.