A Hybrid Deep Learning Approach for Enhancing the Lorentzian Curve Fit Algorithm for Schumann Resonance

Abstract

Schumann Resonance (SR) represents a set of very weak signals formed by the electromagnetic wave propagation along the Earth-ionosphere cavity. Traditionally, SR studies have focused on extracting only their first mode using the average frequency spectrum over long segments. However, advances in sensor technology and digitalization now allow for the analysis of multiple modes, which can significantly enhance our understanding and characterization of SR signals. Even in this context, standard techniques, such as Lorentzian Curve Fit (LCF), often fail to retain the characteristics of the signal when the multiple modes are included, especially under harsh weather conditions. This limitation arises from the mathematical challenges of fitting procedures over noisy, highly variable signals, which become more pronounced when extracting more than three modes. In this paper, we put forward a solution to this problem. Its novelty stems from a Deep Learn-ing (DL) methodology based on three different autoencoder strategies and two learning schemes to denoise signals, handle multiple SR modes simultaneously, while adapt to their non-stationarity and variability. The results demonstrate that the DL model achieves high performance even when fitting a Lorentzian function is impossible, by effectively extracting multiple SR modes under challenging conditions. This outcome provides compelling evidence for further investigation into the use of DL data reduction techniques for analyzing SR signals.