Deep Randomized Feed-forward Networks Based Prediction of Human Joint Angles Using Wearable Inertial Measurement Unit: Performance Comparison

Abstract

Active upper limb assistive robots can improve patients' quality of life with limb impairment and facilitate the rehabilitation of patients in need. To provide intuitive and simultaneous active assistance, continuous joint angle prediction of the upper limb is crucial for promoting the interactive control between the robot and subjects. The model-free approach is gradually becoming mainstream in predicting joint angles, especially those based on deep neural networks. However, the current applied approach can not provide competitive predictive performance, fast computation speed, and learning efficiency. This paper implements random vector functional link networks (RVFL) and extreme learning machine networks (ELM) for the continuous joint angle prediction using only data from wearable inertial measurement units (IMU). The input features are five joint angles of the upper limb in the time domain, derived from the forearm attached IMU and upper arm attached IMU. Five joint angles are fed into the models to predict every joint angle. Multiple RVFL networks (Shallow RVFL, deep RVFL, and deep ensemble RVFL) and ELM networks (Shallow ELM, deep ELM, and deep ensemble ELM) were evaluated over a comprehensive experimental framework. On the one hand, the results show that the prediction performance of RVFL approaches is better than the variant of ELM networks, which are precise for practical robot-assisted application scenarios. On the other hand, the fast computation of RVFL networks offers significant practicability to the real-time control of assistive robots.