DEPS Abstract

Guest >> Sign In

DIRECTED ENERGY PROFESSIONAL SOCIETY

Abstract: 25-Symp-141

UNCLASSIFIED, PUBLIC RELEASE

Application of AI for HEL Beam Control

This paper is intended to summarize the work performed at Spacecraft Research and Design Center (SRDC) at NPS on the application of AI to simplify and improve performance of HEL beam control and focus of future research. The focus of the research in the past has been on two areas.

First is the applications of AI for HEL automatic targeting, aimpoint selection and maintenance. Motivation for this research is that current manual operation of target detection, classification and aimpoint selection is too slow for multiple targets based on current HEL systems. AI algorithms developed under DEJTO project provide automatic HEL targeting, aimpoint selection and maintenance to expedite the process. The accuracy of aimpoint selection by AI was impressive. We have developed 500,000 synthetic images with varying seeing conditions for six UAV models. We also collected over 10,000 laboratory images using HBCRT. The algorithms were evaluated end to end on a UAV trajectory both on computer simulation and implementation on HBCRT with real UAV trajectory. The pose estimation algorithm that was developed under this project is under implementation at NWSC DD on HELIT tracker. The accuracy of aimpoint selection by AI was impressive. However, there are several questions regarding AI applications, such as trustworthiness and accuracy of AI aimpoint prediction, accuracy of prediction for real targets using models developed using synthetic data. and lack of adequate labeled military data sets. Our current research focus is on these areas.

The second area is the application of AI to predict air turbulence from target image and enhancement of target image. Motivation for this research is that we can simplify HEL system and improve performance if from target image we could predict turbulence resulting in simplification of adaptive optics and enhance target image to improve target classification and aimpoint selection. For research, pristine target images were aberrated for a given turbulence. Resnet 15 AI model was trained with aberrated image and turbulence in terms of Zernike polynomials. The results show that AI provides good prediction for weak turbulence. The accuracy decreases for high turbulence. The current focus of research is the use of AI to predict high turbulence.

For target image enhancement, several techniques were compared, blind deconvolution, U-net and DeblurGAN. DeblurGAN provided the best results. Current deep learning models for image aberration compensation and wavefront error prediction are trained separately. We have developed a deep learning training method for concurrent training of image aberration compensation and wavefront error prediction models. The purpose of this concurrent training is to improve the performance of both image aberration compensation and wavefront error prediction models.

UNCLASSIFIED, PUBLIC RELEASE