Investigating the Impact of IMU Sensor Quantity on Locomotion Recognition Performance Using Neural Networks for...

Area of Science:

• Biomedical Engineering
• Robotics
• Machine Learning

Background:

• Powered prostheses offer enhanced mobility for individuals with lower-limb loss compared to passive devices.
• Accurate recognition of locomotion activities is crucial for effective active assistance in prosthetics.
• Minimizing sensor count while maintaining recognition accuracy is a key challenge.

Purpose of the Study:

• To investigate the impact of inertial measurement unit (IMU) sensor quantity on locomotion recognition accuracy.
• To evaluate different neural network (NN) architectures for activity recognition in powered prostheses.
• To identify optimal sensor configurations and NN models for prosthetic locomotion recognition.

Main Methods:

• Utilized the ENABL3S dataset for training and evaluating locomotion recognition models.
• Compared various neural network architectures, including HAR-DeepConvLG.
• Assessed the influence of varying numbers of IMU sensors on recognition performance.

Main Results:

• The HAR-DeepConvLG model demonstrated superior performance, achieving the highest F1 scores (0.73-0.85) across tested sensor quantities.
• Model performance was positively correlated with the number of IMUs used.
• Accurate recognition of level-walking activities was identified as a factor that could further improve overall scores.

Conclusions:

• Neural network architectures and sensor quantity significantly impact locomotion recognition in powered prostheses.
• The HAR-DeepConvLG model shows promise for efficient and accurate activity recognition.
• Optimizing sensor placement and leveraging advanced NN models can enhance prosthetic functionality and user mobility.