ARCHITECTURE OF AN EMBEDDED COMPUTATIONAL STRUCTURE FOR PREDICTIVE ESTIMATION OF UAV DISTURBANCES BASED ON A RECURRENT FUZZY NEURAL NETWORK

Abstract

The presented work focuses on the creation of an intelligent and robust UAV control architecture capable of ensuring stable operation under conditions of time latency, noisy navigation data, and dynamic object uncertainty. The scientific novelty of the research lies in the development and implementation of a recurrent self-evolving neuro-fuzzy network (RSEFNN), integrated into the adaptive switching mode controller (ASMC) circuit for online identification and compensation of non-stationary external disturbances. The key feature of the developed network is the mechanism of autonomous structural adaptation, which allows the system to independently determine the optimal number of fuzzy rules depending on the complexity of the environment, automatically generating new nodes according to the novelty criterion or removing redundant elements through the pruning algorithm. This provides flexible management of computational intensity: the network does not use excessive resources in stable flight modes, but instantly increases power when entering turbulence zones. The introduction of recurrent connections into the network architecture gave the system the properties of dynamic memory, which allows taking into account the history of disturbances without the use of bulky computational structures. The analysis of computational complexity confirmed the high efficiency of the algorithm: the linear dependence of the complexity and the use of SIMD optimization allow maintaining a control frequency of up to 1000 Hz on modern microcontrollers. An important aspect of the software implementation is ensuring determinism of execution time (WCET) through the use of static memory pools, which eliminates the risks of RAM fragmentation. Experimental verification on a UAV digital twin in storm wind conditions confirmed that the use of RSEFNN as a dominant component reduces the positioning error (SS-MAE) by 69.4% compared to standard methods, ensuring precision homing accuracy while maintaining system robustness.