基于深度学习的FAST主动反射面优化研究

Active reflector is used in the Five-hundred-meter Aperture Spherical radio Telescope(FAST). 2225 variable nodes are included in the active reflector, which has 650~720 nodes in the effective aperture of 300m.The large diameter of the aperture and the large amount of nodes bring huge challenges to the real time precision measurement of the reflector of such scale. For current technique status, great difficulties need to be overcome for real time control at the level of several seconds. Programs such as modified Finite-Element-Method model have been arranged. In this project, Deep Learning method in the field of Artificial Intelligence is used to formulate the Neural Networks Models for the active reflector of FAST.The actual measurement data are pre-processed and classified for the time ticks WITH measurement values. The data are then used as training samples and testing data to verify and optimize the model parameters of the Neural Networks Models. The Neural Networks Models are then used to achieve the data forecasting for the time ticks WITHOUT measurement values, and to achieve the optimization of the control of the active reflector of FAST. The project will be essential supplementary and backup of the current works of FAST project.

500米口径大射电望远镜FAST采用主动反射面结构。其主动反射面包括2225个可变节点，300m的有效观测口径内包含650～720个节点。口径大，节点多，给反射面的实时、高精度测量带来巨大挑战。目前的技术条件下，要达到秒级实时性存在不小困难。在现有的状态下，已有工作计划采用有限元模型修正的方法弥补这种局限。本课题计划采用人工智能领域的深度学习模型方法，对有测量数值的时间点的准确测量数据进行预处理和分类，并将其作为训练样本和检验数据，经过深度学习形成神经网络模型，并通过检验优化神经网络模型，从而实现对缺乏测量数值的时间点的有关参数的预测，最终实现对FAST主动反射面控制优化的目标。本课题的开展可以作为现有工作的重要补充和备份。

本项目的背景为FAST主动反射面节点精确调节。主要内容为，通过可以获得的反射面节点位置测量数据，作为神经网络及深度学习的准确参考值，形成深度学习模型. 项目为复杂工况下的反射面节点控制提供新的方法。所实现的硬件和软件系统，对于今后类似望远镜的反射面精确控制和具有参考价值。