Submit
To enhance the adaptive learning, self-organization, and fault tolerance capabilities of gait pattern recognition in intelligent knee prostheses, an online walking pattern recognition method based on the convolutional neural networks (CNN)-long short term memory (LSTM) model is proposed. Five test subjects wore the intelligent knee prostheses and performed four walking modes: level walking, uphill walking, downhill walking, and stair descent. The preprocessed gait data were fed into four neural network models: CNN, LSTM, CNN-LSTM, and CNN-bidirectional LSTM. Through hyperparameter tuning, the recognition accuracy of these models was compared. Real-time indicator, gait recognition delay, was also measured. Experimental results showed each model had its strengths and weaknesses. Overall, the CNN-LSTM model achieved the best recognition performance with accuracy rates of: level walking 89%±2.5%, uphill 72.8%±3.2%, downhill 71%±3.2%, and stair descent 96%±2.5%. When switching from level walking to downhill, gait recognition delay was 51.7%±15.6%, and vice versa it was 75.8%±11.5%; when switching from level walking to stair descent, gait recognition delay was 47.1%±17.1%, and vice versa it was 38.6%±10.5%. In summary, the application of the CNN-LSTM model for walking pattern recognition in unilateral intelligent knee prostheses is feasible, with accuracy and real-time performance meeting the control requirements of the prostheses.
Keywords: Neural network, walking pattern, gait recognition, recognition delay rate
