Optimized EEG Multi-Noise Removal and Compression: Deploying a PbP-QLP Enhanced Autoencoder on STM32 Microcontroller

Abstract

Electroencephalograms (EEGs) are effective and patient-friendly for diagnosing, monitoring, and preventing mental disorders. However, due to their low voltage, EEG signals often contain noise that obscures critical features, risking misdiagnosis. Current denoising methods typically address one or two noise types and struggle with memory limitations on edge devices. To overcome these challenges, we introduce a quantization-based compressed denoising autoencoder (DAE) model using a PbP-QLP, a low-rank approximation (LRA) technique, for multi-noise removal (15 types, including power-line, baseline wander, ocular, muscle artifacts, and combinations) in EEGs on low-memory edge devices. Our compression technique reduces the model size from 8 to 1.51 MB, achieving 81% weight compression with minimal loss.