A HYBRID MD-CNN UAV FAULT PREDICTION METHOD WITH FLIGHT DATA

This paper proposed a Mahalanobis distance - convolutional neural network (MD-CNN) fault prediction method using real flight data from unmanned aerial vehicles (UAVs). The development of UAV technology has led to its wide application in various industries. Early detection of undetectable anomalies in UAVs is essential for improving their reliability, ensuring flight safety, and avoiding economic losses. UAV flight data, as a typical highdimensional large sample dataset, requires data-driven fault prediction methods to predict faults and improve reliability. The proposed method used Mahalanobis distance discrimination method to detect partial fault data of UAVs, and the detection results are served as the input of one-dimensional convolutional neural network (1D-CNN) to predict UAV faults. Experimenting with real UAV flight data, the proposed method achieved accuracy, precision, recall and F1-score of 97.30%, 96.78%, 96.99% and 96.88% for UAV fault prediction, respectively. The experimental results indicated that the established MD-CNN method showed significant advantages over traditional anomaly detection algorithms based on the One-Class Support Vector Machine (OCSVM) method and machine learning fault diagnosis algorithms based on SVC, Gaussian Naive Bayes, KNN, and Decision Tree, which could accurately predict faults before the occurrence of UAV faults.