In daily rescue missions, Unmanned Aerial Vehicles (UAVs) are vital. Finding propeller damage in UAVs is a critical first step in guaranteeing safe operation and endurance. Because a faulty UAV might endanger people nearby, coexisting with humans and UAVs brings additional concerns. In Addition, by promptly and effectively identifying these problems, the risk to human life can be minimized to the greatest extent possible by providing neighbors with an early warning. The main purpose of this study is to develop an inspection model to classify healthy and broken propellers through audio sound generated by the drone propeller. In addition, different audio data augmentation mechanisms are used to increase the data volume size. We have applied two approaches to train the model: passive learning and active learning. Furthermore, four different CNN-based architectures are used to develop the acoustic inspection models which were trained with three different datasets to find the model’s performances and their training accuracy and loss. In the passive deep learning approaches, the ResNet-50 inspection model has attained the best accuracy which is 88.54% by using the entire dataset. In contrast, the ResNet-50 inspection model obtained 85.86% model accuracy in the active deep learning approach, the highest accuracy among the iterations.
AI-Driven Quadrocopter Propeller Acoustic Health Monitoring based on Deep Learning
2024-11-06
935793 byte
Conference paper
Electronic Resource
English
TIBKAT | 2011
|British Library Online Contents | 2009
Optimization-based iterative learning for precise quadrocopter trajectory tracking
British Library Online Contents | 2012
|