A Novel BILSTM-CNN Method for Pattern Recognition in Real Time Under Triple Physical Loads in Lower Extremity Exoskeleton

Abstract

Pattern recognition is of great importance in compliant control of lower extremity exoskeleton. We propose a novel method based on the BILSTM-CNN model for pattern recognition in real time under triple physical loads on different terrains. BILSTM is skilled in dealing with temporal series data while CNN is proficient in coping with spatial series data. Five patterns include level ground walking (LW), stair ascending (SA), stair descending (SD), ramp ascending (RA), and ramp descending (RD). Their accuracies in multigrade loads of 0kg, 20kg, and 40kg reach 98.81%, 98.24%, and 97.04% respectively. Moreover, compared with the universal methods of LSTM, CNN, and BP, the hybrid method has competitive advantage in evaluation indexes of accuracy, precision, recall, and F1 score. In addition to five steady patterns, eight pattern transitions are identified between two neighboring states, such as LW to SA, LW to SD, LW to RA, LW to RD, SA to LW, SD to LW, RA to LW, and RD to LW. For pattern transitions, prediction time (Pre-T) of next pattern in multilevel loads of 0kg, 20kg, and 40kg are 190-620 ms, 180-420 ms, and 50-90 ms respectively prior to the step into that pattern. Pattern transition time (Aug-In) in multilevel loads of 0kg, 20kg, and 40kg are 50-190 ms, 30-310 ms, and 0-280 ms respectively before Pre-T of next pattern. Eventually, the experimental results indicate the proposed method has excellent performance on pattern recognition and pattern transition.