A deep learning strategy for EMG-based joint position prediction in hip exoskeleton assistive robots. (May 2022)

• Record Type: Journal Article
• Authors: Foroutannia, Ali; Akbarzadeh-T, Mohammad-R; Akbarzadeh, Alireza
• Is Part Of: Biomedical signal processing and control. Volume 75(2022)
• Publication Date: 2022-05
• Subjects: Lower limb exoskeletons -- Assistive robotics -- Electromyography -- Joint position prediction -- Deep learning -- Long short-term memory network; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.103557
• Languages: English
• ISSNs: 1746-8094

Abstract

Graphical abstract: Highlights: An in-depth learning strategy for predicting the hip joint position of an exoskeleton robot using EMG signals is proposed. The EMG, FSR, load cell, waist angle, and hip joint position when walking at five speeds for seven healthy individuals are recorded. A CNN is provided for position estimation and an LSTM network for predicting the hip joint position of an exoskeleton robot. Trained networks are considered in an impedance control loop for joint trajectory estimation and prediction. The proposed method reduces exoskeleton impedance control error and human effort. Abstract: The exoskeleton robots for the lower limb can help meet the necessary hip joint force to rehabilitate people with movement disorders. This paper proposes a deep learning strategy using electromyography (EMG) signals to predict the human hip joint position and to determine the exoskeleton robot's necessary auxiliary force at the next step. For this purpose, the EMG signals, force-sensing resistor (FSR), load cell, waist angle with inertial measurement unit (IMU), and hip joint position when walking at speeds of 0.3, 0.4, 0.5, 0.6, and 0.85 m/s for seven healthy individuals are recorded. After the initial preprocessing and extraction of the EMG features, deep neural networks are used in two stages. First, the position of the hip joints is estimated by a convolutional neural network. Second, a long short-term memory (LSTM) network is presented to predict the future hip position. The trained networks are then placed in an impedance control loop, which controls the robot online towards the predicted joint trajectory. Experiments have also been performed to evaluate the accuracy and robustness of the proposed position estimation and prediction algorithm. Results indicate that using the predicted position by the proposed deep strategy reduces the controller error and leads to better synchronization of the exoskeleton robot with its human user, reducing/supporting the required human effort in walking.