A Deep Learning Control Strategy of IMU-Based Joint Angle Estimation for Hip Power-Assisted Swimming Exoskeleton

Wearable exoskeleton techniques are becoming mature and widely used in many areas. However, the biggest challenge lies in that the control system should recognize and follow the wearer's motion correctly and quickly. In this study, we propose a deep learning control strategy using inertial measurement units (IMUs) for hip power-assisted swimming exoskeleton. The control strategy includes two steps: Step 1: the swimming stroke is recognized by a deep convolutional neural and bidirectional long short-term memory network (DCNN-BiLSTM) and Step 2: the hip joint angles are estimated with BiLSTM network belonging to the recognized motion to predict the hip trajectory. The dataset of motion recognition and estimation of four swimming strokes is collected by placing IMUs on swimmers' back and thighs. We conduct offline and online testing of control strategy for accuracy and robustness validation. During offline testing, we achieve an accuracy of more than 96% of motion recognition and root mean square error (RMSE) less than 1.2° of hip joint angle estimation, outperforming 2.76% of accuracy and 0.09° of RMSE compared with those of extreme learning machine (ELM) or conventional neural network and gate recurrent unit (CNN-GRU). During online testing, the pretrained networks are transplanted into a Raspberry Pi 4B and achieve 8.47 ms for conducting one motion recognition and 6.72 ms for one hip joint angle estimation on average, which are far less than 300 ms of delayed sensations between the action of exoskeleton and human, while keeping a satisfying recognition accuracy as well. The experimental results show that the accuracy and robustness of the proposed control strategy are stable and feasible for application to exoskeletons.