Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

An LSTM Feature Imitation Network for Hand Movement Recognition from sEMG Signals

Chuheng Wu, S. Farokh Atashzar, Mohammad M. Ghassemi, Tuka Alhanai

TL;DR

The paper addresses data scarcity in sEMG-based hand movement recognition by introducing a Feature Imitation Network (FIN) that pre-trains an LSTM to imitate four explicit time-domain features, enabling interpretable temporal feature learning within a 300 ms window. The FIN is followed by a feature-augmented 2D-CNN classifier; pre-training on these features and subsequent fine-tuning improves cross-subject transfer and data efficiency, outperforming SVM baselines and end-to-end CNNs trained on raw signals. Results show $R^2$ values above $0.96$ for feature imitation and accuracy gains of up to several tens of percent when using FIN-enabled transfer learning, with robustness to limited data and near-future (up to 300 ms) prediction. This approach offers practical, data-efficient sEMG representations suitable for robust hand movement recognition and prosthetic control, with future work exploring other FINs and broader transfer scenarios.

Abstract

Surface Electromyography (sEMG) is a non-invasive signal that is used in the recognition of hand movement patterns, the diagnosis of diseases, and the robust control of prostheses. Despite the remarkable success of recent end-to-end Deep Learning approaches, they are still limited by the need for large amounts of labeled data. To alleviate the requirement for big data, we propose utilizing a feature-imitating network (FIN) for closed-form temporal feature learning over a 300ms signal window on Ninapro DB2, and applying it to the task of 17 hand movement recognition. We implement a lightweight LSTM-FIN network to imitate four standard temporal features (entropy, root mean square, variance, simple square integral). We observed that the LSTM-FIN network can achieve up to 99\% R2 accuracy in feature reconstruction and 80\% accuracy in hand movement recognition. Our results also showed that the model can be robustly applied for both within- and cross-subject movement recognition, as well as simulated low-latency environments. Overall, our work demonstrates the potential of the FIN modeling paradigm in data-scarce scenarios for sEMG signal processing.

An LSTM Feature Imitation Network for Hand Movement Recognition from sEMG Signals

TL;DR

The paper addresses data scarcity in sEMG-based hand movement recognition by introducing a Feature Imitation Network (FIN) that pre-trains an LSTM to imitate four explicit time-domain features, enabling interpretable temporal feature learning within a 300 ms window. The FIN is followed by a feature-augmented 2D-CNN classifier; pre-training on these features and subsequent fine-tuning improves cross-subject transfer and data efficiency, outperforming SVM baselines and end-to-end CNNs trained on raw signals. Results show values above for feature imitation and accuracy gains of up to several tens of percent when using FIN-enabled transfer learning, with robustness to limited data and near-future (up to 300 ms) prediction. This approach offers practical, data-efficient sEMG representations suitable for robust hand movement recognition and prosthetic control, with future work exploring other FINs and broader transfer scenarios.

Abstract

Surface Electromyography (sEMG) is a non-invasive signal that is used in the recognition of hand movement patterns, the diagnosis of diseases, and the robust control of prostheses. Despite the remarkable success of recent end-to-end Deep Learning approaches, they are still limited by the need for large amounts of labeled data. To alleviate the requirement for big data, we propose utilizing a feature-imitating network (FIN) for closed-form temporal feature learning over a 300ms signal window on Ninapro DB2, and applying it to the task of 17 hand movement recognition. We implement a lightweight LSTM-FIN network to imitate four standard temporal features (entropy, root mean square, variance, simple square integral). We observed that the LSTM-FIN network can achieve up to 99\% R2 accuracy in feature reconstruction and 80\% accuracy in hand movement recognition. Our results also showed that the model can be robustly applied for both within- and cross-subject movement recognition, as well as simulated low-latency environments. Overall, our work demonstrates the potential of the FIN modeling paradigm in data-scarce scenarios for sEMG signal processing.
Paper Structure (15 sections, 3 figures, 3 tables)

This paper contains 15 sections, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Data and Methodology
  3. Data Acquisition and Processing
  4. Features
  5. Model Architecture
  6. Feature Imitating LSTM and Augmentation
  7. Downstream Classifier
  8. Experiments
  9. Exp. $\#$1: Feature Imitating Experiment
  10. Exp. $\#$2: Downstream CNN Training
  11. Exp. #3: Downstream Fine-tuning Experiments
  12. Model Training Time
  13. Exp. #4: Data Efficiency Validation
  14. Exp. #5: Forward-in-time Classification
  15. Conclusion

Figures (3)

  • Figure 1: Bi-directional Feature Imitating LSTM Neural Network; 3-layer structure, each layer has 600 forward-backward connected LSTM cells. Each LSTM-FIN learns to generate one of four features (RMS, VAR, ENT, SSI).
  • Figure 2: Downstream neural network architecture for recognition, mapping from augmented features to classifications
  • Figure 3: LSTM-FIN+CNN-II Model Performance Across Data Sizes and Future Predictions. (Top) Illustrates the mean accuracy given various proportions of training data. (Bottom) Illustrates the mean accuracy given variations in time to future event (100 samples = 50ms).