Leveraging reduced-order models for state estimation using deep learning
Leveraging reduced-order models for state estimation using deep learning
复制标题
DOI：
10.1017/jfm.2020.409
复制
发表时间：
2020-08-25
影响因子：
3.7
通讯作者：
Goza, Andres
中科院分区：
文献类型：
--
作者：
Nair, Nirmal J.;Goza, Andres
状态估计是分析物理机制和实现流体流动实时控制的关键。一种常见的估计方法是将传感器测量值与由降阶模型（ROM）控制的简化状态相关联。(When如果需要，可以通过ROM恢复完整状态。）在这一类别中的当前方法几乎总是使用线性模型来将传感器数据与简化状态相关联，这通常导致对传感器位置的限制，并且在表示测量与简化状态之间的一般非线性关系方面具有固有的限制。我们提出了一种替代方法，利用神经网络架构来学习这种非线性关系。神经网络是这个估计问题的自然选择，因为减少状态传感器测量关系的物理解释很少是明显的。所提出的估计框架是不可知的ROM所采用的，并可以被纳入任何选择的ROM上得到的线性子空间（例如适当的正交分解）或非线性流形。所提出的方法被证明是一个二维模型问题的分离流绕平板，并发现优于常见的线性估计的替代品。
State estimation is key to both analysing physical mechanisms and enabling real-time control of fluid flows. A common estimation approach is to relate sensor measurements to a reduced state governed by a reduced-order model (ROM). (When desired, the full state can be recovered via the ROM.) Current methods in this category nearly always use a linear model to relate the sensor data to the reduced state, which often leads to restrictions on sensor locations and has inherent limitations in representing the generally nonlinear relationship between the measurements and reduced state. We propose an alternative methodology whereby a neural network architecture is used to learn this nonlinear relationship. A neural network is a natural choice for this estimation problem, as a physical interpretation of the reduced state-sensor measurement relationship is rarely obvious. The proposed estimation framework is agnostic to the ROM employed, and can be incorporated into any choice of ROMs derived on a linear subspace (e.g. proper orthogonal decomposition) or a nonlinear manifold. The proposed approach is demonstrated on a two-dimensional model problem of separated flow around a flat plate, and is found to outperform common linear estimation alternatives.