Joint Torque Estimation Using sEMG and Deep Neural Network
- PDF / 3,309,848 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 16 Downloads / 197 Views
ORIGINAL ARTICLE
Joint Torque Estimation Using sEMG and Deep Neural Network Harin Kim1 · Hyeonjun Park1 · Sangheum Lee1 · Donghan Kim1 Received: 10 October 2019 / Revised: 27 March 2020 / Accepted: 19 June 2020 © The Korean Institute of Electrical Engineers 2020
Abstract With the aid of various physical and biological sensors, research is actively being conducted to understand the intention of wearer’s motions through parameters such as joint torque. sEMG signals can be measured faster than physical sensors, which are often used in the field of behavioral intent identification studies. However, electrodes must be placed in the correct positions, and due to the high volume of noise, professional knowledge and accurate hardware design are required. In this paper, a system is constructed to improve the sEMG signal measurement environment by producing small multichannel sEMG modules. In addition, deep neural network supervised learning algorithms are implemented to estimate the torque using only the sEMG signal. Based on this, we analyze the organization of algorithms, the processing of the sEMG data, and how the number of channels affects learning. The optimal deep natural network model selected by the analysis is implanted to embedded after learning. The implanted model performs a portable real-time torque optimization (PRTE) according to the sEMG signal entered. In this paper, we study the deep natural network algorithm for estimating sEMG hardware and torque, and how it is implanted into a portable embedded system for use in estimating real-time motion intent. The proposed deep natural network algorithm and the usefulness of the PRTE system are verified through experiments. Keywords Electromyogram · Multi channels surface-Electromyography (sEMG) · Upper limb · Joint torque estimation · Deep neural network · Regression · PRTE system
1 Introduction Recently, exoskeleton robot systems have been developed to support human movement for various purposes in fields such as military, recycling, industrial, and disaster remediation. Typical examples are Cyberdyne’s HAL, which uses a combination of sEMG and physical sensors to assist with lower muscle strength, and Ford’s Eksovest, which relieves the burden of workers by providing support for upper extremity muscles based on physical sensors [1]. Wearable robots work together by directly * Donghan Kim [email protected] Harin Kim [email protected] Hyeonjun Park [email protected] Sangheum Lee [email protected] 1
Department of Electrical Engineering, Kyung Hee University, Yongin‑si, Republic of Korea
contacting each other and performing specific target tasks, such as supporting the muscle strength of the wearer. It is important to have a natural interaction between the wearer and wearable robot to cope dynamically in various environments. When the exoskeleton robot assists in human behavior, it is necessary to know the required force, in real time, to identify the load and force that the user is responsible for. This is because exoskeleton can assist a user’s movement
Data Loading...
