Two-Stage Iterative Finite-Memory Neural Network Identification for Unmanned Aerial Vehicles

Recently, a system identification technique called FiMos-TA (Kang et al., 2022) has been introduced for the accurate identification of unmanned aerial vehicles (UAVs). This approach offers impressive performances, such as robustness and accuracy against disturbances and error accumulation, by utilizing a finite-memory-based training scheme. However, a major limitation of FiMos-TA (Kang et al., 2022) is that it requires inverse matrix operations on large matrices to obtain training gains, which severely affects its real-time implementation performance. Moreover, it relies on the amount of variations in the initial weights and the measurement model in finite-memory is insufficiently adaptable to changing conditions. Therefore, we propose a new approach called the two-stage iterative finite-memory neural network (TSIFNN) identification strategy for UAVs that overcomes all the limitations of FiMos-TA (Kang et al., 2022), ensuring not only robustness and accuracy but also real-time performance. We demonstrate the real-time performance, robustness, and accuracy of the proposed TSIFNN identification through a UAV experiment.