Speaker
Description
Seismic and environmental noise, particularly Newtonian noise, constitute a fundamental
limitation for the Einstein Telescope, motivating the development of
advanced, data-driven approaches for noise mitigation to improve the sensitivity
and operational stability of its interferometric systems. Real-time compensation
of Newtonian noise can improve sensitivity of the sensor and allow rapid
triggering of multi-messenger observations when a gravitational wave signal is
detected.
A neural network is trained to reconstruct the waveform of a target sensor
based on measurements from neighboring nodes. The method considers an array
of N three-axis accelerometers, sampled at rates of up to 4,000 Hz. Correlation
analysis is used to identify suitable data preprocessing techniques and to determine
the optimal input window length for the neural network. The network
and data preprocessing is aimed to be implemented on an FPGA to achieve
real-time signal processing. For this purpose, tools such as hls4ml or similar
frameworks will be used.
The approach is first evaluated using publicly available seismic datasets.
This is followed by experimental validation using an FPGA-based sensor network
with micro-electromechanical systems (MEMS) accelerometers, which collect
seismic data and transmit it to a central FPGA for real-time prediction
of target sensor measurements. Data compression techniques will be used to
optimize communication between FPGA nodes.