Replay-Based Incremental Learning Framework for Gesture Recognition Overcoming the Time-Varying Characteristics of...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Surface electromyography (sEMG) based gesture recognition faces challenges like electrode displacement and signal variability over time, impacting cross-time application reliability.
• Existing methods struggle with catastrophic forgetting and adapting to non-synchronous data features in dynamic environments.

Purpose of the Study:

• To propose a robust incremental learning framework for sEMG gesture recognition that addresses signal instability and catastrophic forgetting.
• To enhance the accuracy and efficiency of sEMG-based gesture recognition for practical, real-world applications.

Main Methods:

• Developed an incremental learning framework utilizing densely connected convolutional networks (DenseNet) to process non-synchronous sEMG data.
• Implemented a replay dataset strategy storing data across different time spans for joint model training, mitigating catastrophic forgetting.
• Employed the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm for efficient selection of representative samples to update the replay dataset.

Main Results:

• Achieved an average recognition rate of 96.5% across eight subjects after multiple incremental updates, outperforming cross-day analysis.
• Demonstrated superior performance with DBSCAN for sample selection, reaching a 93.7% recognition rate using fewer samples compared to conventional methods.
• Showcased incremental learning's advantage over full dataset training, with a nearly 1% improvement in recognition rate, reduced training time, and lower iteration costs.
• Attained an 88.6% average recognition rate for incremental learning of action classes, enabling flexible addition of new gestures.

Conclusions:

• The proposed incremental learning framework effectively handles sEMG signal instability and catastrophic forgetting, offering a significant improvement in gesture recognition accuracy and robustness.
• The framework's efficiency in training time and model updating, coupled with its ability to incorporate new action classes, makes it highly suitable for practical sEMG-based applications.
• This approach enhances the adaptability and application value of action pattern recognition technology using sEMG signals in dynamic scenarios.