Data-driven diagnosis of multicopter thrust fault using supervised learning with inertial sensors

This study proposes a data-driven fault diagnosis for multicopter unmanned aerial vehicles that uses the principal direction vector of inertial measurement unit (IMU) sensor signals calculated by principal component analysis. The main idea comes from the fact that a normal sphere-shaped distribution of the sensor data changes to a specific elliptical shape under a certain thrust fault situation. The fault diagnosis is based on classification and regression using supervised learning with the gyroscope and accelerometer datasets of an IMU. We analyze the performance of the proposed approach by depending on different learning algorithms. To verify the diagnostic performance, ground experiments with a hexacopter on the gimbaled jig are performed for various cases of damaged propellers. Then, the applicability of the proposed data-driven fault diagnosis is confirmed by analyzing the accuracy of the fault’s location and degree.