Distributed State Estimation With Deep Neural Networks for Uncertain Nonlinear Systems Under Event-Triggered Communication

This work explores the distributed state estimation problem for an uncertain, nonlinear, and continuous-time system. Given a sensor network, each agent is assigned deep neural network (DNN) that used to approximate system's dynamics. Each updates weights of their DNN through multiple timescale approach, i.e., outer layer are updated online with Lyapunov-based gradient descent update law, inner concurrently using supervised learning strategy. To promote efficient use resources, observer uses event-triggered communication. A nonsmooth Lyapunov analysis demonstrates achieves uniformly ultimately bounded reconstruction. simulation example five-agent estimating two-link robotic manipulator tracking desired trajectory provided validate result showcase performance improvements afforded by DNNs.