Deep Learning-Based Robust Actuator Fault Detection and Isolation Scheme for Highly Redundant Multirotor UAVs

Abstract

This article presents a novel approach for detecting and isolating faulty actuators in highly redundant Multirotor UAVs using cascaded Deep Neural Network (DNN) models. The proposed Fault Detection and Isolation (FDI) framework combines Long Short-Term Memory (LSTM)-based fault detection and faulty actuator locator models to achieve real-time monitoring. The study focuses on a Hexadecarotor multirotor UAV equipped with sixteen rotors. To tackle the complexity of FDI resulting from redundancy, a partitioning technique is introduced based on system dynamics. The proposed FDI scheme is composed of a region classifier model responsible for detecting faults and fault locator models that precisely determine the location of the failed actuator. Extensive training and testing of the models demonstrate high accuracy, with the regional classifier model achieving 98.97% accuracy and the fault locator model achieving 99.107% accuracy. Furthermore, the scheme was integrated into the flight control system of the UAV, before being tested via both real-time monitoring in the simulation environment and analysis of recorded real flight data. The models exhibit remarkable performance in detecting and localizing injected faults. Therefore, using DNN models and the partitioning technique, this research offers a promising method for accurately detecting and isolating faulty actuators, thereby improving the overall performance and dependability of highly redundant Multirotor UAVs in various operational scenarios.

Keywords

• FDI
• Multirotor UAV
• UAM
• LSTM
• RNN
• DNN