Unsupervised noise reductions for gravitational reference sensors or accelerometers based on Noise2Noise method

Zhilan Yang, Haoyue Zhang, Peng Xu, Ziren LuoarXiv 2023 · DOI · arXiv

TL;DR

Researchers have successfully adapted the "Noise2Noise" deep learning framework to denoise high-precision gravitational sensors used in space missions. By using clever sub-sampling techniques, the method removes complex environmental noise from satellite accelerometers without ever needing "clean" training data, significantly outperforming traditional filters.

What Problem Does This Paper Address?

High-precision inertial sensors (accelerometers) are the heart of gravity-mapping satellites like GRACE-FO and gravitational wave detectors like Taiji-1. However, these sensors operate in a "noisy" neighborhood. They are subject to complex interference from temperature fluctuations, magnetic fields, and the satellite platform itself.

Traditional noise reduction—like Kalman or Butterworth filters—often struggles because the noise is "colored" (not just simple white noise) and overlaps with the actual scientific signals. While deep learning offers a solution, standard "supervised" models require a "clean" version of the data to learn from. In space, we only have the noisy measurements; there is no "ground truth" sensor providing a perfect signal for comparison. This paper tackles the challenge of denoising signals when you only have a single, noisy sample to work with.

Methods

The authors utilize an unsupervised approach called Noise2Noise (N2N). Originally designed for image processing, N2N operates on a simple but powerful principle: if you have two different noisy realizations of the same underlying signal, a network can learn to map one to the other. In doing so, it learns the underlying signal because the noise (being random and independent) cannot be predicted.

To make this work for 1D time-series sensor data, the authors introduced two novel sub-samplers:

  1. Odd-Even Sub-sampler: For general signals, the data is split into odd and even points to create the "noisy pair."
  2. Periodic Sub-sampler: For signals that repeat (like during specific satellite maneuvers), the team used data from different cycles to form the training pairs.

They tested two neural network architectures: a U-Net (famous for its "skip connections" that preserve detail) and a Convolutional Autoencoder (CAE). A key innovation was adding a reconstruction layer (fully connected layers) to the CAE, which helped the model converge faster and better handle low-frequency noise.

Datasets

The study validated the method using three distinct data sources:

  • Simulation Data: A synthetic square wave combined with both Gaussian and "colored" noise to establish a baseline.
  • Taiji-1 Data: Real acceleration data from China’s first gravitational wave detection technology satellite, specifically during calibration maneuvers.
  • GRACE-FO Data: Level-1A acceleration data from the Gravity Recovery and Climate Experiment Follow-On mission.

Key Insights and Ideas

  • Learning from Noise: The most profound insight is that the network doesn't need to know what "clean" looks like. As long as the noise in the input and the noise in the target are independent, the network's loss function naturally gravitates toward the signal's mean value.
  • Temporal Independence: By splitting a single time series into odd and even points, the authors effectively created "independent" noise samples from a single measurement stream, bypassing the need for multiple sensors.
  • The Power of Periodicity: For satellite maneuvers, the periodic sub-sampler allows the model to "see" the same signal across different time windows, making it exceptionally robust against transient glitches.
  • Preserving the Peak: Unlike traditional low-pass filters that might "smear" or dampen important resonance peaks in the data, the N2N method suppressed the surrounding noise floor while keeping the critical signal peaks sharp and clear.

Key Results

The N2N method consistently beat out traditional signal processing:

  • Superior SNR: In simulations, the N2N method achieved a Signal-to-Noise Ratio (SNR) of 23.56, compared to just 16.58 for a Butterworth filter and 12.64 for Wavelet denoising.
  • Taiji-1 Improvements: When applied to real Taiji-1 data, the method made resonance peaks in the Amplitude Spectral Density (ASD) significantly clearer, which is vital for calibrating the satellite's center of mass.
  • GRACE-FO Precision: For GRACE-FO data, the N2N method suppressed high-frequency and impulsive noise so effectively that the residuals (the difference between the denoised data and the official Level-1B processed data) were at the $10^{-10}$ level, demonstrating "space-grade" accuracy.

Limitations and Future Work

While highly effective, the authors noted that convolutional layers can suffer from "edge effects" where the beginning and end of a data segment aren't denoised as well. They mitigated this using a sliding window approach with 50% overlap, but it adds computational overhead.

Future work will focus on developing even more complex N2N processing schemes and exploring how this unsupervised deep learning can be integrated into the standard data pipelines for future gravitational wave missions (like LISA or Taiji), where the signal-to-noise requirements are incredibly stringent.

Summary updated

Abstract

Onboard electrostatic suspension inertial sensors are important applications for gravity satellites and space gravitational wave detection missions, and it is important to suppress noise in the measurement signal. Due to the complex coupling between the working space environment and the satellite platform, the process of noise generation is extremely complex, and traditional noise modeling and subtraction methods have certain limitations. With the development of deep learning, applying it to high-precision inertial sensors to improve the signal-to-noise ratio is a practically meaningful task. Since there is a single noise sample and unknown true value in the measured data in orbit, odd-even sub-samplers and periodic sub-samplers are designed to process general signals and periodic signals, and adds reconstruction layers consisting of fully connected layers to the model. Experimental analysis and comparison are conducted based on simulation data, GRACE-FO acceleration data and Taiji-1 acceleration data. The results show that the deep learning method is superior to traditional data smoothing processing sol
BibTeX
@misc{yang2023unsupervised,
  author = {Zhilan Yang and Haoyue Zhang and Peng Xu and Ziren Luo},
  title = {{Unsupervised noise reductions for gravitational reference sensors or accelerometers based on Noise2Noise method}},
  year = {2023},
  eprint = {2305.06735},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2305.06735},
}
Figure