Data Augmentation to Address Various Rotation Errors of Wearable Sensors for Robust Pre-impact Fall Detection

Area of Science:

• Biomedical Engineering
• Sensor Technology
• Machine Learning

Background:

• Wearable sensors are vital for fall detection, but rotation errors due to attachment and movement degrade performance.
• Robust pre-impact fall detection is critical for timely intervention and injury prevention.

Purpose of the Study:

• To investigate the efficacy of data augmentation in mitigating sensor rotation errors for fall detection.
• To compare uniform and normal data augmentation strategies against a non-augmented model.

Main Methods:

• Developed two augmented models (uniform and normal) and a non-augmented baseline.
• Evaluated models on original and validation datasets with varying pitch, roll, and compound roll and pitch (CRP) rotation errors (15°, 30°, 45°).
• Utilized five-fold cross-validation to assess model accuracy and lead time.

Main Results:

• Augmented models maintained high accuracy (>98.5%) on original data and improved performance by over 6% on rotated data.
• CRP errors impacted accuracy most, followed by pitch and roll; the normal augmentation strategy excelled in typical error ranges.
• Augmented models demonstrated significant improvements in lead time on the validation dataset.

Conclusions:

• Data augmentation effectively addresses sensor rotation errors in wearable fall detection systems.
• Augmented models, particularly the normal strategy, show strong potential for practical, robust pre-impact fall detection and injury prevention applications.