Radio Telescope Surface Measurement via Deep Learning

Abstract

This paper proposes a new method for accurately measuring the surface deformation of radio telescope antennas based on deep learning. A deep convolutional neural network is used to predict surface deformations by mapping the near-field intensity of the antenna, instead of relying entirely on a physical model. The proposed method could offer precise measurement of surface deformations in real time with only a single image of near-field intensity pattern. To optimize the deep learning model, a preliminary U-net based deep convolutional neural network (DCNN) model was developed based on a large data set generated by an approximate physical model, a partial differential equation (PDE). The network parameters were then fine-tuned using transfer learning with a small data set obtained by high precision numerical simulation. During this process, the fine-tuning layers that achieved optimal performance for the U-net network was studied. The final results show that the proposed method significantly improves the accuracy of antenna surface deformation recovery. Additionally, singular value decomposition (SVD) technology is employed to denoise the intensity image, which facilitates the application of the proposed method to actual deformation measurement.