UAV Fault and Anomaly Detection Using Autoencoders
Raju Dhakal, Carly Bosma, Prachi Chaudhary, Laxima Niure Kandel
Abstract
The popularity of Uncrewed Aerial Vehicles (UAVs) is on the rise, but these complex systems are susceptible to faults and anomalies impacting their safety and performance. To deal with emergency situations, it is necessary to monitor the status of these aircraft and report any anomalies or faults. Therefore, it is of great significance to study the anomaly detection method for UAV systems. In this study, unsupervised neural network models called Autoencoders (AE) and Variational Autoencoders (VAE) are utilized to detect UAV faults and anomalies. The key idea is to train autoencoders to learn the normal data and, after training, use them to identify the abnormal data by observing the magnitude of the reconstruction error. This serves as both an indicator of anomalies during inference and a cost function in training. Our results from publicly available real UAV sensor data called ALFA (Air Lab Failure and Anomaly) verify that the VAE-based method can effectively detect faults and anomalies with an average accuracy of 95.6%.