Characterizing seismic isolation using convolutional neural networks and Wiener filters

TL;DR

This study compares neural networks and Wiener filters in characterizing seismic motion transfer in a gravitational wave detector testbed, revealing strengths and limitations of each method across frequency ranges.

Contribution

It introduces a neural network approach with time-dilation layers for seismic data analysis and compares it to Wiener filters, optimizing large filter computations.

Findings

SISO Wiener filters outperform MISO filters due to low coherence.

Neural network with Huber loss performs slightly worse than Wiener filters.

FFT-based neural network loss outperforms Wiener filters at low amplitudes and in certain frequency regions.

Abstract

We investigate seismic motion propagation through a passively isolated mechanical system, using Wiener filters and convolutional neural networks with time-dilation layers. The goal of this study was to explore the capabilities of neural networks and Wiener filters in characterizing a mechanical system from the measurements. The mechanical system used is a testbed facility for technology development for current and future gravitational wave detectors, "VATIGrav", currently being commissioned at University of Hamburg. It consists of a large vacuum chamber mounted on four active vibration isolators with an optical table inside, mounted on four passive vibration isolators. In this paper we have used seismic data recorded on the ground and on the optical table inside the chamber. The data were divided in 6 hours for training and another 6 hours for validation, focusing on inferring 150-second stretches of time series of table motion from the ground motion in the frequency range from $0.1~\mathrm{Hz}$ to about $50~\mathrm{Hz}$. We compare the performance of a neural network with FTT-based loss function and with Huber loss function to single-input, single-output (SISO) and multiple-input, single-output (MISO) Wiener filters. To be able to compute very large MISO Wiener filters (with 15,000 taps) we have optimized the calculations exploiting block-Toeplitz structure of the matrix in Wiener-Hopf equations. We find that for the given task SISO Wiener filters outperform MISO Wiener filters, mostly due to low coherence between different motion axes. Neural network trained with Huber loss performs slightly worse than Wiener filters. Neural network with FFT-based loss outperforms Wiener filters in some frequency regions, particularly with low amplitudes and reduced coherence, while it tends to slightly underestimate the peaks, where Wiener filters perform better.