Acoustic Anomaly Detection for Propeller Crack Defects: A Drone-Based Surrogate Dataset for UAM

초록

The imminent commercialization of Urban Air Mobility (UAM) necessitates robust AI-enabled maintenance to ensure passenger and public safety. A nondestructive acoustic method is presented for detecting cracks in UAM propellers using a surrogate dataset recorded from drone propellers. The dataset, comprising a total of 5,760 recording files, contains recordings of normal operating sounds and two abnormal categories (ripped and broken) captured while varying microphone-propeller angles and throttle levels. Short-time Fourier transform (STFT) preprocessing and an autoencoder-based reconstruction model enable defect detection when training on normal acoustic data only. In angle-specific tests, the approach achieved ROC AUC up to 0.978 and F1 up to 1.0. In throttle power-specific tests reached an ROC AUC up to 0.9785. The dataset and methodology provide a foundation for future UAM maintenance applications.

키워드