Pattern transition recognition based on transfer learning for exoskeleton across different terrains.

Area of Science:

• Robotics and Human-Computer Interaction
• Machine Learning and Pattern Recognition

Background:

• Human motion intention detection is crucial for advanced wearable robots.
• Recognizing transitions between locomotion modes (e.g., walking, stair climbing) is challenging, especially under varying physical loads and terrains.

Purpose of the Study:

• To develop and evaluate a novel transfer learning method for accurate and early recognition of locomotion mode pattern transitions.
• To improve the performance of wearable robots in adapting to different terrains and physical conditions.

Main Methods:

• A transfer learning approach combining Temporal Convolutional Network (TCN) with Spatial Attention (SA) was developed for pattern transition recognition.
• The method was tested on recognizing transitions among eight locomotion modes across five dynamic modes under triple physical loads.
• Performance was evaluated based on accuracy and prediction time (Pre-T) compared to other models like ResNet and LSTM.

Main Results:

• The TCN-Spatial Attention (TCN-SA) transfer learning method achieved high accuracy, reaching up to 98.21%.
• Early prediction of the next locomotion mode was achieved, with prediction times ranging from 120-600 ms before the actual transition.
• The proportion of prediction time in a gait cycle (Pre-T/GC) was significantly reduced, indicating efficient prediction.

Conclusions:

• The proposed TCN-SA transfer learning method effectively detects locomotion pattern transitions, enabling earlier predictions.
• This early detection allows wearable robots, such as exoskeletons, to navigate between adjacent terrains more smoothly and efficiently.
• The approach demonstrates superior performance compared to existing methods in prediction time and accuracy.