Unsupervised Noise Reductions for Gravitational Reference Sensors or Accelerometers Based on the Noise2Noise Method

Area of Science:

• Space physics
• Sensor technology
• Signal processing

Background:

• Onboard electrostatic suspension inertial sensors are crucial for gravity satellites and gravitational-wave detection.
• Complex space environments create intricate noise patterns, limiting traditional noise modeling and subtraction methods.
• Improving the signal-to-noise ratio in high-precision inertial sensors is essential for mission success.

Purpose of the Study:

• To apply deep learning techniques to enhance signal quality in onboard electrostatic suspension inertial sensors.
• To address limitations of traditional noise reduction methods in complex space environments.
• To develop a novel deep learning approach for noise suppression in satellite-based inertial sensing.

Main Methods:

• Designed odd-even and periodic sub-samplers for general and periodic signals.
• Integrated reconstruction layers with fully connected networks into the deep learning model.
• Validated the model using simulated data, GRACE-FO, and Taiji-1 acceleration data.

Main Results:

• The deep learning method demonstrated superior performance compared to traditional data smoothing.
• The proposed approach effectively suppresses noise in inertial sensor measurements.
• Experimental results confirmed the efficacy of the deep learning model on real-world satellite data.

Conclusions:

• Deep learning offers a powerful solution for noise reduction in high-precision inertial sensors for space applications.
• The developed sub-sampling and reconstruction techniques improve the signal-to-noise ratio in challenging orbital conditions.
• This work advances the potential of deep learning in space exploration and fundamental physics missions.