TL;DR
This paper introduces a novel incremental learning framework for sEMG-based gesture recognition that tackles the challenge of signal instability over time due to factors like electrode drift. By combining DenseNet with a replay-based approach and a smart sample selection method (DBSCAN), the framework achieves high accuracy (96.5%) and significantly reduces training time, making it more practical for real-world applications.
What Problem Does This Paper Address?
Surface electromyography (sEMG) signals are promising for gesture recognition, enabling applications in prosthetics, human-computer interaction, and rehabilitation. However, a major hurdle for long-term, real-world use is the inherent instability of sEMG signals. This instability arises from factors like electrode displacement and changes in muscle fatigue over time, leading to a significant drop in recognition accuracy when models trained on one day are used on data from another. Traditional methods often require retraining on large, updated datasets, which is computationally expensive and impractical. This paper aims to address this "cross-time" recognition problem by developing an incremental learning approach that can adapt to these time-varying signal characteristics without catastrophic forgetting of previously learned gestures.
Methods
The core of the proposed solution is a replay-based incremental learning framework built upon a DenseNet (Densely Connected Convolutional Network).
- DenseNet for Feature Extraction: DenseNet is chosen for its ability to effectively extract features from the one-dimensional sEMG data. Its dense connectivity pattern helps mitigate gradient vanishing and degradation issues, allowing for deeper networks and improved performance.
- Replay-Based Incremental Learning: To combat "catastrophic forgetting" (where a model forgets old knowledge when learning new data), the framework employs a replay strategy. This involves storing a subset of previously seen data (the "replay dataset") and retraining the model using both the new incremental data and this stored old data. This allows the model to continuously learn new patterns while retaining its knowledge of past ones.
- DBSCAN for Replay Dataset Update: A key innovation is the use of the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm for selecting representative samples to update the replay dataset. Instead of random or uniform sampling, DBSCAN identifies dense clusters of similar data points, effectively selecting more representative samples and filtering out outliers. This is crucial for maintaining the quality of the replay data and improving recognition accuracy with fewer samples.
- Class Incremental Learning: The framework also extends to class incremental learning, allowing new gesture classes to be added to the model over time, further enhancing its adaptability.
Datasets
The study utilized a multi-day sEMG dataset collected from eight healthy adult male volunteers over 12 days. Each subject performed seven common gestures (wrist extension, wrist flexion, wrist radial deviation, wrist ulnar deviation, wrist supination, wrist pronation, and clenched fist). The data was collected using a four-channel sEMG acquisition device at a sampling frequency of 1000 Hz. This dataset simulates the real-world scenario of varying signal quality and characteristics over extended periods.
Key Insights and Ideas
- Time-Varying sEMG is a Major Challenge: The study strongly validates that electrode displacement and signal drift significantly degrade performance in cross-time sEMG recognition, showing accuracy drops of up to 35.6% compared to single-day analysis.
- Replay Learning is Effective for sEMG: The replay-based incremental learning approach proves to be a robust solution for maintaining performance over time, outperforming simple cross-day analysis.
- DBSCAN is a Smart Way to Curate Replay Data: Using DBSCAN for sample selection in the replay dataset significantly improves recognition rates compared to random or uniform selection, especially when dealing with fewer samples. This highlights the importance of data quality in incremental learning.
- Incremental Learning Offers Practical Advantages: Beyond accuracy, the incremental approach drastically reduces training time and computational costs compared to retraining on full datasets, making it far more suitable for real-time or continuously updating systems.
- Class Incremental Learning Adds Flexibility: The ability to incrementally add new gesture classes makes the system adaptable to evolving user needs or new gesture sets without complete retraining.
Key Results
- High Overall Accuracy: The proposed replay-based incremental learning framework achieved an impressive average recognition rate of 96.5% across eight subjects after multiple increments.
- Superior to Cross-Day Analysis: This significantly outperforms a simple cross-day analysis, which showed an average recognition rate of 64.2%.
- DBSCAN Performance: The DBSCAN-based sample selection for the replay dataset achieved a recognition rate of 93.7% with fewer samples, outperforming other methods.
- Efficiency Gains: Incremental learning training time was consistently around 10 seconds, compared to 64 seconds for full dataset training (after nine increments), representing a substantial speedup.
- Accuracy Improvement over Full Training: The incremental learning approach even showed a slight improvement of nearly 1% in recognition accuracy compared to training with the complete dataset.
- Class Incremental Performance: For class incremental learning, the framework achieved an average recognition rate of 88.6% after adding new classes, demonstrating its capability to expand its gesture repertoire.
Limitations and Future Work
- Time Span of Data: The study acknowledges that the experimental design had a relatively short time span (12 days). The authors suggest that future research should investigate longer time spans to further refine dynamic weighting strategies for the replay dataset.
- Dynamic Weighting: While the study used a fixed weight for new data, it suggests that a dynamic weighting approach, adjusting weights based on the validity of data over longer time spans, could further improve performance.
- Class Incremental Performance Degradation: As observed in other studies, class incremental learning did lead to a slight decrease in overall performance compared to the initial model. Future work could focus on mitigating this degradation more effectively.
- Generalizability: While promising, further validation on larger and more diverse datasets, including different populations and a wider range of gestures, would be beneficial.
Summary updated