End-to-End Deep Fault-Tolerant Control

PUBLISHED ON IEEE/ASME TRANSACTIONS MECHATRONICS, DOI: 10.1109/TMECH.2021.3100150. Ideally, accurate sensor measurements are needed to achieve a good performance in the closed-loop control of mechatronic systems. As consequence, faults will prevent system from working correctly, unless fault-tolerant (FTC) architecture is adopted. model-based FTC algorithms for nonlinear systems often challenging design, this paper focuses on new method presence faults, based deep learning. The considered approach replaces phases fault detection and isolation controller design with single recurrent neural network, which has value past given time window as input, current values variables output. This end-to-end applied composed spherical inverted pendulum, whose configuration changed via reaction wheels, turn actuated by electric motors. simulation experimental results show that proposed can handle abrupt occurring link position/velocity sensors. provided supplementary material includes video real-world experiments software source code.