A Surrogate Tiny Machine Learning Model of Variational Autoencoder For Real-Time Baseline Correction of Magnetometer Data

• Published in: 2024 International Conference on Computing, Networking and Communications (ICNC) pp. 579 - 583
• Main Authors: Siddique, Talha, Mahmud, MD Shaad
• Format: Conference Proceeding
• Language: English
• Published: IEEE 19.02.2024
• Subjects: Computational modeling; Data Denoising; Edge Machine Learning; Geomagnetic Data; Magnetometers; Performance evaluation; Probabilistic logic; Satellites; TinyML; Uncertainty; Variational Autoencoder; Weather forecasting
• ISSN: 2473-7585
• DOI: 10.1109/ICNC59896.2024.10556237

Magnetometers play a vital role in geophysics and space weather prediction applications by collecting terrestrial magnetic field data. They monitor solar-induced geomagnetic disturbances, providing essential insights into predicting space weather effects on technologies such as satellites, power grids, and communication networks. However, this data often contains inherent background noise, necessitating accurate baseline correction methods. Traditional correction techniques are robust but computationally demanding and unsuitable for real-time applications. Recent progress has investigated the utilization of Tiny Machine Learning (TinyML) to process magnetometer data in real time, especially when resources are limited. However, these edge-based ML solutions often lack the robustness of more computationally intensive probabilistic models, such as Variational Autoencoders (VAEs). This paper introduces a TinyML-VAE surrogate model designed for real-time magnetometer baseline correction. The surrogate model approximates an implemented VAE's performance while operating within the constrained resources of an edge device. The new model retains the VAE's uncertainty quantification capabilities by leveraging surrogate modeling techniques, ensuring robustness. Experimental out- comes have been displayed, illustrating a comparison between the performance of the TinyML-VAE and the benchmark established by the standard VAE.