Least-order representation of control-oriented flow estimation exemplified for the fluidic pinball

Abstract Deep learning has been widely utilized to accurately estimate the flow state from sparse sensor measurements. Yet there is still a lack of understanding actual dimension this type regression problem. In study, we propose an Autoencoder (AE) based estimation method tackle control-oriented, sensor-based This encodes input information into few-nodes bottleneck, then decodes compressed estimated state. network stands for least order representation input-output relationship. We choose fluidic pinball as benchmark problem which contains multiple-input and multiple-output. The rotations three cylinders in generates rich physics. Consequently, strong non-linearity arises between measurements output For fair comparison Neural Network (DNN) also employed highlight influence encoding-decoding process AE. results show that when periodic, vector can be effectively encoded five-dimensional subspace without significant loss accuracy. However, encoding will eliminate partial forced becomes chaotic, resulting lower accuracy than DNN.