Enhancing Upper Limb Exoskeletons Using Sensor-Based Deep Learning Torque Prediction and PID Control

• Published in: Sensors (Basel, Switzerland) Vol. 25; no. 11; p. 3528
• Main Authors: Shakeriaski, Farshad, Mohammadian, Masoud
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 03.06.2025; MDPI
• Subjects: Accuracy; Artificial intelligence; Biomechanics; Brain research; Control systems; Control theory; Datasets; Deep learning; deep learning models; Design; Elbow; Machine learning; Neural networks; Performance evaluation; proportional–integral–derivative control algorithm; Quality of life; Rehabilitation; Robotics; Robots; Sensors; Stroke; stroke rehabilitation; Systems stability; torque estimation and prediction; upper limb assistive exoskeleton robot elbow orthosis
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s25113528

Upper limb assistive exoskeletons help stroke patients by assisting arm movement in impaired individuals. However, effective control of these systems to help stroke survivors is a complex task. In this paper, a novel approach is proposed to enhance the control of upper limb assistive exoskeletons by using torque estimation and prediction in a proportional–integral–derivative (PID) controller loop to more optimally integrate the torque of the exoskeleton robot, which aims to eliminate system uncertainties. First, a model for torque estimation from Electromyography (EMG) signals and a predictive torque model for the upper limb exoskeleton robot for the elbow are trained. The trained data consisted of two-dimensional high-density surface EMG (HD-sEMG) signals to record myoelectric activity from five upper limb muscles (biceps brachii, triceps brachii, anconeus, brachioradialis, and pronator teres) during voluntary isometric contractions for twelve healthy subjects performing four different isometric tasks (supination/pronation and elbow flexion/extension) for one minute each, which were trained on long short-term memory (LSTM), bidirectional LSTM (BLSTM), and gated recurrent units (GRU) deep neural network models. These models estimate and predict torque requirements. Finally, the estimated and predicted torque from the trained network is used online as input to a PID control loop and robot dynamic, which aims to control the robot optimally. The results showed that using the proposed method creates a strong and innovative approach to greater independence and rehabilitation improvement.