A method of classification decision based on multi-BiLSTMs for physical loads hierarchy.

Area of Science:

• Biomechanics
• Wearable Technology
• Machine Learning

Background:

• Human gait is significantly altered by physical loads.
• Accurate load recognition is crucial for controlling wearable assistive devices.
• Existing methods require further refinement for multi-load scenarios.

Purpose of the Study:

• To develop and assess an algorithm for classifying different physical load levels using wearable inertial measurement units (IMUs).
• To evaluate the effectiveness of load classification as a strategy for multi-load recognition in exoskeleton control.
• To compare the proposed algorithm's performance against traditional methods like LSTM and BiLSTM.

Main Methods:

• Collected gait data from 10 adults walking with varying backpack loads (0, 15, 25 kg) using lower limb sensors.
• Developed a classification decision method based on multiple bidirectional long short-term memory networks (multi-BiLSTMs).
• Real-time data acquisition and classification of gait patterns under different load conditions.

Main Results:

• The proposed multi-BiLSTMs algorithm achieved a classification accuracy of 94.1%.
• The F-score ranged from 0.935 to 0.952, indicating high precision and recall.
• The multi-BiLSTMs method demonstrated superior performance compared to standard LSTM and BiLSTM models for load classification.

Conclusions:

• The developed algorithm effectively classifies physical load levels from human gait data.
• This approach offers a promising strategy for real-time load quantification and adaptive exoskeleton control.
• Potential applications include medical rehabilitation and occupational safety, enhancing human-exoskeleton interaction.