کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

STA-EEGNet: Spatial-Temporal Attention Enhanced EEGNet Architecture with Adaptive Model Compression for Efficient Brain-Computer Interfaces

نویسندگان :

Mohsen Avesta¹ Maedeh Avesta² Asma Yousefian Baboukani³ Bashir Najafabadian⁴ Mohaddeseh Behjati⁵

1- Faculty of Electrical Engineering, Shahid Beheshti University, Tehran, Iran 2- Department of Biomedical Engineering, Najafabad Branch, Islamic Azad University 3- Department of Biomedical Engineering, Sheikh Bahaei University 4- Department of Biomedical Engineering, Science and Research Branch, Islamic Azad University, Tehran 5- 5Interventional Cardiology Research Center, Cardiovascular Research Institute, Isfahan University of Medical Sciences, Isfahan, Iran

کلمات کلیدی :

EEGNet،brain-computer interface،spatial temporal attention،model compression،adaptive quantization،channel pruning،knowledge distillation،edge computing،neural architecture search

چکیده :

This paper presents STA-EEGNet, a novel spatialtemporal attention enhanced architecture for EEG signal classification, integrated with a comprehensive adaptive model compression framework. The proposed spatial-temporal attention mechanism mathematically models the interdependencies between EEG channels and temporal dynamics through a hierarchical attention formulation. We introduce Adaptive Channel Selection Quantization (ACSQ), Progressive Structured Pruning (PSP), and Multi-level Teacher Assistant Distillation (MTAD) as novel compression techniques specifically optimized for EEG processing. Theoretical analysis proves the convergence properties of our attention mechanisms with a convergence rate of O(1/T ). Extensive experiments on four benchmark EEG datasets (BCI Competition IV 2a, PhysioNet MI, DEAP, TUH EEG Corpus) demonstrate that STA-EEGNet achieves state-of-the-art accuracy of 75.6%, 89.7%, 83.5%, and 86.8% respectively, while the compression framework reduces model size by 8.7× and inference latency by 12.3× with only 1.0% accuracy degradation. The compressed model achieves real-time performance (53.5 FPS) on ESP32-S3 microcontrollers with 86KB memory footprint, enabling practical deployment in wearable BCI systems.