DEPS Abstract

Guest >> Sign In

DIRECTED ENERGY PROFESSIONAL SOCIETY

Abstract: 24-Symp-021

UNCLASSIFIED, PUBLIC RELEASE

In-situ machine-learning-based high-rate atmospheric path characterization sensing signal forecasting for beam control and laser communication applications

Forecasting in machine learning (ML) refers to the process of predicting future values expected several timesteps ahead based on historical signal time evolution and trends. It involves developing models that capture the underlying signal patterns in the time series data to make predictions about future signal values.
Time ahead forecasting of signal future values may play an important role in atmospheric EO systems, including HEL DE and laser communication systems. The performance of these systems is strongly affected by the continuously changing atmospheric conditions along the propagation path occurring at time scales of a few seconds or less in operation with stationary targets, and even faster in operations with moving targets. These performance variations can be forecasted, and potentially mitigated, when atmospheric turbulence conditions are evaluated and forecast in real-time (in-situ).
Time series forecasting of the power-in-the-bucket (PIB) signal (metric) used for coherent beam combining (CBC) in fiber array-based HEL DE systems prevents the negative impact of occasional deep signal declines below levels acceptable for CBC system operation and enables in-situ selection of time windows optimal for HEL beam firing. Similarly, the negative impact of deep received [e.g., power-in-the-fiber (PIF)] signal fades in laser communications systems can be mitigated (e.g., by in-situ adaptation of the data transmission rate) if these fades can be forecasted several timesteps ahead.
In this presentation we discuss the ongoing development of ML-based forecasting of future values of both the PIB metric and the atmospheric refractive index structure parameter Cn2 time series measured during several atmospheric field trials at the UD 7 km test range under diverse turbulence conditions. Results demonstrate that the ML techniques developed can provide high-accuracy (MSE <5%, 10 datapoints, or 2 msec ahead) forecasting of PIB sensing values captured by the sensors at high (5 kHz, or 0.2 ms) acquisition rates. For forecasting of Cn2 values, we used time series of Cn2 predictions computed by TurbNet DNN models based on in-situ sensing of short-exposure intensity scintillation patterns at rates of 1.5 sec per a single measurement, as described in Ref. [1]. We show that the developed ML technique can provide accurate forecasting of Cn2 values up to 15 seconds ahead.
[1] E. Polnau, D. Hettiarachchi and M. Vorontsov “Electro-Optical Sensors for Atmospheric Turbulence Strength Characterization with Embedded Edge AI Processing of Scintillation Patterns,” Photonics 9, 789, (2022)

UNCLASSIFIED, PUBLIC RELEASE