Multi-Fault Diagnosis in UAV Electrical Power Systems Using a Dual-Stage CNN

Abstract

Fault diagnosis is an important condition for reliable and safe UAV electrical power systems in multiple fault condition, which can be achieved with accurate and timely diagnosis. To overcome this challenge, the authors in this study suggest a dual-stage convolutional neural network (CNN) approach to multi-fault diagnosis in UAV electrical power systems. The proposed architecture combines two complementary classification steps: one based on a fault detection model that classifies normal and abnormal operating conditions, and another based on a fault isolation model that classifies faulty sensors or system components. The framework was built and assessed with the NASA ADAPT dataset, which consists of 122 fault scenarios and 292,169 samples gathered from the operation of UAV electrical power systems. Data preprocessing, hyperparameter optimization, comparative evaluation and confusion matrix analysis were used to evaluate the effectiveness of the proposed approach. Results have shown high Precision, Recall, F1-score, ROC-AUC and Matthew's correlation coefficient with 99.10% and 99.30% classification accuracy for fault detection and fault isolation respectively. Comparative experiments also demonstrated superior performance of the proposed CNN-based approach to SVM, Random Forest, XGBoost, ANN and LSTM models under the considered conditions. Furthermore, the model stability, interpretation and the role of the most important CNN parts in its working were investigated through five-fold cross validation, SHAP-based feature analysis, and an ablation study. The results show that using the proposed dual-stage CNN, multi-fault diagnosis is an effective method and can serve as a foundation for intelligent health monitoring and reliability improvement of electrical power systems in UAVs.