Vibration data-driven anomaly detection in UAVs: A deep learning approach

Abstract

Unmanned Aerial Vehicles (UAVs) are employed for diverse applications, including aerial surveillance and package delivery. However, the occurrence of faults, especially propeller failures, poses significant risks to safe and efficient operations. Detecting such faults at an early stage is critical to avoiding catastrophic outcomes and ensuring the reliability and lifespan of UAVs. To address this crucial need, this study proposes a novel approach for monitoring vibration signals using a wavelet scattering long short-term memory (LSTM) autoencoder network. The LSTM autoencoder can learn temporal patterns from input signals, whereas wavelet scattering can capture the dynamics and interactions of various frequency components of signals. First, a deliberate modification was made to one of the propeller blades of the DJI M600 multi-rotor UAV to deliberately induce vibration. The proposed network was then evaluated on the acquired vibration signal using the MTi-G-700 IMU. The results showed that warning signals were generated from all axes before failures occurred. Notably, the earliest warnings were obtained from y-axis data within 100 s, while the first warning from z-axis data was recognized 130 s later. The failure occurred at roughly 280 s. The experimental findings indicate that the proposed method can accurately detect anomalies that could potentially lead to failure.