Real-time Flow Estimation via Limited Sensor Data Using Closed-Loop Neural Networks Observers

This work presents results on the application of neural network observers trained in closed-loop. A recurrent training approach is proposed in a setup similar to the classic discrete-time Luenberger observer. Results for a controlled Lorenz system are presented, where the states are estimated from a set of real-time measurements. A confined cylinder flow is also investigated, where velocity signals at the wake are estimated from a set of wall pressure probes, enabling feedback control using neural network controllers.