A CNN-Based Method for Intent Recognition Using Inertial Measurement Units and Intelligent Lower Limb Prosthesis

• Published in: IEEE transactions on neural systems and rehabilitation engineering Vol. 27; no. 5; pp. 1032 - 1042
• Main Authors: Su, Ben-Yue, Wang, Jie, Liu, Shuang-Qing, Sheng, Min, Jiang, Jing, Xiang, Kui
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.05.2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adult; Algorithms; Amputation; Amputees; Ankle; Artificial Limbs; Artificial neural networks; Ascent; Biomechanical Phenomena; Calibration; convolutional neural networks (CNN); Data analysis; Descent; Electromyography; Empirical analysis; Humans; inertial measurement unit (IMU); Inertial platforms; Intent recognition; Intention; Knee; Learning algorithms; Leg; Legged locomotion; Locomotion; lower limb prosthesis; Machine Learning; Male; Mapping; Measurement units; Mechanical sensors; Neural networks; Neural Networks, Computer; Pattern recognition; Prostheses; Prosthesis Design; Prosthetics; Recognition, Psychology; Steady state; swing phase; Trans-femoral amputees; Walking
• ISSN: 1534-4320; 1558-0210; 1558-0210
• DOI: 10.1109/TNSRE.2019.2909585

Powered intelligent lower limb prosthesis can actuate the knee and ankle joints, allowing transfemoral amputees to perform seamless transitions between locomotion states with the help of an intent recognition system. However, prior intent recognition studies often installed multiple sensors on the prosthesis, and they employed machine learning techniques to analyze time-series data with empirical features. We alternatively propose a novel method for training an intent recognition system that provides natural transitions between level walk, stair ascent / descent, and ramp ascent / descent. Since the transition between two neighboring states is driven by motion intent, we aim to explore the mapping between the motion state of a healthy leg and an amputee's motion intent before the upcoming transition of the prosthesis. We use inertial measurement units (IMUs) and put them on the healthy leg of lower limb amputees for monitoring its locomotion state. We analyze IMU data within the early swing phase of the healthy leg, and feed data into a convolutional neural network (CNN) to learn the feature mapping without expert participation. The proposed method can predict the motion intent of both unilateral amputees and the able-bodied, and help to adaptively calibrate the control strategy for actuating powered intelligent prosthesis in advance. The experimental results show that the recognition accuracy can reach a high level (94.15% for the able-bodied, 89.23% for amputees) on 13 classes of motion intent, containing five steady states on different terrains as well as eight transitional states among the steady states.