Deep generative models with data augmentation to learn robust representations of movement intention for powered leg...

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Intent recognition in powered leg prostheses traditionally relies on expert rules or extensive labeled data.
• Collecting sufficient training data from amputee populations is challenging and time-consuming.
• Covariate shift can negatively impact prosthesis controller accuracy if movement intention representations are not robust.

Purpose of the Study:

• To develop and evaluate data augmentation and deep neural network techniques for robust movement intention recognition.
• To generate realistic synthetic sensor data for training prosthesis controllers.
• To improve the generalizability of intent recognition models in wearable robotics.

Main Methods:

• Utilized deep neural networks and data augmentation strategies to learn robust representations of movement intention.
• Collected sensor data from four amputee subjects over three days for offline analysis.
• Generated synthetic sensor data to supplement real-world training and testing datasets.

Main Results:

• The developed approach successfully produced realistic synthetic sensor data.
• Error rates were reduced when training and testing on different days and with different users.
• Demonstrated the effectiveness of data augmentation in enhancing the robustness of intent recognition models.

Conclusions:

• Data augmentation combined with deep learning offers an effective and generalizable strategy for wearable robotics sensor data.
• This approach challenges the notion that rehabilitation robotics benefits minimally from deep learning due to data limitations.
• The findings suggest a pathway to improve the performance and adaptability of powered prosthetic devices.