Adaptive estimation model: Robust full-state prediction through sparse observations with variable layout and quantity

Abstract

Recovering the full-state from limited observation data is crucial because it provides a reliable reference for active control. Advances in deep learning technology further enable building robust and high-precision estimators. In this work, we proposed a robust estimator capable of adapting to the number and layout of observation points, enabling precise reconstruction of the pressure distribution from sparse observations on the hydrofoil surface. The ability for adaptive and accurate reconstruction stems from carefully designed steps of randomly masking inputs in the training stage and Transformer models with enhanced feature extraction capabilities. An impact analysis was conducted to assess the influence of the number of observation points on reconstruction performance. Furthermore, tests to evaluate the performance under conditions with missing observation points were also conducted. These assessments validated the robust reconstruction performance of the proposed model. Noise tests further highlighted the strengths of the proposed model, but also revealed stability issues that can be improved.