Machine learning based techniques for failure detection and prediction in Unmanned Aerial Vehicle

Unmanned aerial vehicles (UAVs) are aircraft without human pilot on the board. UAVs have two flight mode: Auto and Manual. In Auto mode, UAV follows a predefined path. The path is embedded in the control system of UAV. In manual mode, human operator in ground control station controls the trajectory of the vehicle remotely. UAVs have diverse applications in military as well as civil sectors. UAVs have to operate in a variety of unseen environments. The diverse usage and uncertainty in operational environment demand safe and reliable operation. Timely identification and rectification of faults stand as a crucial requirement for the operation. One of the major cause of UAV failure is engine fault. In this paper we investigate affectiveness of machine learning techniques regarding engine fault detection and prediction. We analyzed the techniques on AirLab Failure and Anomaly (ALFA) Dataset. For fault detection we used Multi-Layer Perceptron, Random Forest, Support Vector Machine, Ada Boost, Gradient Boosting, Logistic Regression and Single Dimensional Convolutional Neural Network. We observed that Random Forest is most effective technique for fault detection with F-1 Score of 0.99. Regarding fault prediction we tried LSTM and GRU based network in different settings. Gated Recurrent Unit performed best with F-1 Score of 0.99 while predicting fault four second ahead of time.