# Sensor Head Temperature Distribution Reconstruction of High-Precision Gravitational Reference Sensors with Machine Learning

## TL;DR

This paper uses machine learning to accurately reconstruct sensor head temperatures in gravitational reference sensors, which is crucial for space-based experiments.

## Contribution

A novel XGBoost-LSTM method is proposed for high-precision temperature reconstruction in constrained sensor environments.

## Key findings

- The XGBoost-LSTM method achieves two orders of magnitude higher precision than conventional interpolation methods.
- The method shows one order of magnitude better performance than BP neural networks and meets Taiji Program requirements.
- The approach demonstrates stability and robustness in both ground-based and on-orbit simulation scenarios.

## Abstract

Temperature fluctuations affect the performance of high-precision gravitational reference sensors. Due to the limited space and the complex interrelations among sensors, it is not feasible to directly measure the temperatures of sensor heads using temperature sensors. Hence, a high-accuracy interpolation method is essential for reconstructing the surface temperature of sensor heads. In this study, we utilized XGBoost-LSTM for sensor head temperature reconstruction, and we analyzed the performance of this method under two simulation scenarios: ground-based and on-orbit. The findings demonstrate that our method achieves a precision that is two orders of magnitude higher than that of conventional interpolation methods and one order of magnitude higher than that of a BP neural network. Additionally, it exhibits remarkable stability and robustness. The reconstruction accuracy of this method meets the requirements for the key payload temperature control precision specified by the Taiji Program, providing data support for subsequent tasks in thermal noise modeling and subtraction.