Scholarlink

<p>Surface electromyography (sEMG) signals provide valuable insights into muscle activity patterns. However, sEMG-based human activity recognition remains challenging due to noise interference and signal crosstalk. This study addresses these challenges by analyzing sEMG data from healthy and pathological subjects, using a public dataset comprising 11 healthy individuals and 11 subjects with knee ailments performing three activities: walking, sitting with leg extension, and standing with leg flexion. To enhance classification accuracy and reliability, we propose a novel approach that integrates variational mode decomposition (VMD) with a hybrid deep learning model. Our method reconstructs the signal by selecting optimal intrinsic mode functions based on signal-to-noise ratio, correlation, and reconstruction error, effectively reducing noise and isolating informative features. The proposed hybrid convolution neural network long short-term memory CNN-LSTM model achieves high classification accuracies: 99.1% for healthy and 98.7% for pathological subjects. Furthermore, we employ explainable AI (XAI) techniques to interpret model predictions. The permuted feature importance (PFI) analysis reveals distinct muscle activation patterns: healthy subjects exhibit balanced reliance on all muscle groups during walking, rectus femoris and vastus medialis are dominant during sitting, while biceps femoris and semitendinosus play a crucial role in standing with flexion. Conversely, pathological subjects rely more on the vastus medialis as a compensatory mechanism due to inhibited activation in the rectus femoris. These findings can inform targeted rehabilitation strategies for musculoskeletal disorders. Additionally, our study contributes to human activity recognition (HAR) by demonstrating how sEMG-based models can distinguish movement patterns in both healthy and pathological populations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive VMD-Based Explainable Hybrid Deep Learning Model for Robust Lower Limb Activity Recognition from sEMG Signals

  • 期刊论文

  • 发表日期：

    2025年11月8日

查看全文
  • Pratibha Tokas,
  • Vijay Bhaskar Semwal,
  • Sweta Jain

摘要

Surface electromyography (sEMG) signals provide valuable insights into muscle activity patterns. However, sEMG-based human activity recognition remains challenging due to noise interference and signal crosstalk. This study addresses these challenges by analyzing sEMG data from healthy and pathological subjects, using a public dataset comprising 11 healthy individuals and 11 subjects with knee ailments performing three activities: walking, sitting with leg extension, and standing with leg flexion. To enhance classification accuracy and reliability, we propose a novel approach that integrates variational mode decomposition (VMD) with a hybrid deep learning model. Our method reconstructs the signal by selecting optimal intrinsic mode functions based on signal-to-noise ratio, correlation, and reconstruction error, effectively reducing noise and isolating informative features. The proposed hybrid convolution neural network long short-term memory CNN-LSTM model achieves high classification accuracies: 99.1% for healthy and 98.7% for pathological subjects. Furthermore, we employ explainable AI (XAI) techniques to interpret model predictions. The permuted feature importance (PFI) analysis reveals distinct muscle activation patterns: healthy subjects exhibit balanced reliance on all muscle groups during walking, rectus femoris and vastus medialis are dominant during sitting, while biceps femoris and semitendinosus play a crucial role in standing with flexion. Conversely, pathological subjects rely more on the vastus medialis as a compensatory mechanism due to inhibited activation in the rectus femoris. These findings can inform targeted rehabilitation strategies for musculoskeletal disorders. Additionally, our study contributes to human activity recognition (HAR) by demonstrating how sEMG-based models can distinguish movement patterns in both healthy and pathological populations.