Advances and Prospects in Motor Intent Prediction Based on Surface Electromyography Signals to Aid Upper Limb Exoskeleton Control

This paper introduces two basic technologies, surface electromyography signals (sEMG) processing and motion intention prediction models. To begin with, the system describes the main directions of signal preprocessing and feature extraction, filtering and denoising, and a multi-source synchronization mechanism, and compares representational dissimilarities between time-frequency domain features and deep learning features. Secondly, two mainstream models: discrete action recognition and continuous motion estimation are explored in this paper. It discusses the benefits of deep architectures like Transformer and multi-stream fusion networks to operate on long-term time-series dependencies and summarizes the three key bottlenecks that the current technology encounters in balancing real-time operations and real-time performance with accuracy, cross-individual flexibility, and resilience in complex settings. Last but not least, in three dimensions of having a multimodal perception fusion, low-weight computing models, and adaptive learning, the objective of the paper is to propose a theoretical reference point in designing a highly accurate and low-latency upper limb exoskeleton control system.