Fixed-Wing Unmanned Aerial Vehicle Rotary Engine Anomaly Detection via Online Digital Twin Methods
Chao-Chung Peng, Yi-Ho Chen
Abstract
Anomaly detection based on data-driven methods is an applicable way to deal with the complex structure of aircraft engine. In this article, certain existing data-driven methods are first introduced for model construction of a fixed-wing unmanned aerial vehicle (UAV) rotary engine. However, due to the prediction transient response and the associated stability not being guaranteed, a hybrid observer/Kalman filter identification (OKID) scheme is proposed. The presented method uses autoregressive model with exogenous inputs (ARX) model for modeling and involves a deadbeat observer design, which can allow model outputs to converge to real output in a theoretical proof. The identified models are seen as the digital twins of a healthy system, which can be taken as a reference to monitor the status of the UAV rotary engine. For comparison study, three data-driven methods, including neural network (NN), fast orthogonal search (FOS), and the proposed OKID hybrid model, are assessed by their model accuracy, stability, and convergence through practical flight data. Experimental results show that the developed method is the best alternative for online fault detection even in the face of limited training data. Moreover, given the real test flight data, the proposed OKID hybrid model can identify the anomaly status and figure out the abnormal part of the fixed-wing rotary engine, which greatly contributes to field managers for maintenance policy decision-making.