Deep Learning Framework for Structural Faults Classification of a Nano Quadcopter

This paper describes a deep learning framework for fault classification of a nano-quadcopter. Based on Crazyflie data, several structural faults are analyzed such as blade damage, weight and inertia imbalance, etc. The Deep Neural Network (DNN) achieved very high accuracy (98.56%) and the fastest inference time (0.92 seconds), though it requires longer training (533.73 seconds). XGBoost provides a good balance with high accuracy (96.17%) and a relatively fast inference time (3.64 seconds). The k-Nearest Neighbors (KNN) model has a swift training time (0.08 seconds) but suffers from a long inference time (58.74 seconds). LightGBM, while having the highest accuracy (98.67%), faces challenges with the longest training (2367.61 seconds) and inference times (145.76 seconds). Nevertheless, robust feature engineering is found to improve the model performance as well as the transparency and explainability of AI systems in safety-critical applications.