Hybrid FOPID and DL on EMG signals for gait phases classification to rehabilitation robot control: A comparative study

摘要

Abstract Rehabilitation after neurological or musculoskeletal impairments requires accurate detection and prediction of gait phases to support motor recovery and ensure safe, adaptive assistance. However, traditional assessment techniques often depend on subjective clinical evaluations, lacking the precision and responsiveness needed for real‐time robotic interventions. To overcome these limitations, we propose a hybrid intelligent rehabilitation system that combines machine learning (ML) models with fractional order proportional‐integral‐derivative (FOPID) controllers for enhanced gait phase classification and trajectory correction. The proposed methodology integrates electromyography (EMG) signals, joint angle kinematics (hip and knee), and foot pressure data of 1000 samples to train and evaluate three machine learning models: random forest (RF), support vector machine (SVM), and long short‐term memory (LSTM). Feature extraction techniques are applied to capture time‐domain and frequency‐domain characteristics of the EMG and biomechanical signals. The initial ML‐based classification identifies eight distinct gait phases. Subsequently, a FOPID controller is employed to correct misclassifications and refine the trajectory of lower‐limb articulations in real time. Experimental evaluations were conducted using a dataset collected from healthy subjects performing continuous walking trials. Results show that the RF‐FOPID hybrid model achieves the best classification performance, with an accuracy of 82% across all gait phases. In terms of trajectory tracking, the integration of the FOPID controller leads to significant reductions in mean absolute error (MAE), with improvements of 39.12% at the hip joint and 40.21% at the knee. In an other hand, the LSTM‐FOPID model showed the highest improvement in error correction, with 50% to 60% reduction in tracking deviations compared with the uncorrected baseline. These findings highlight the effectiveness of combining fractional‐order control with machine learning to improve the precision, robustness, and adaptability of rehabilitation robots. This hybrid approach offers promising applications in personalized gait rehabilitation, allowing for real‐time correction and adaptive assistance tailored to the patient's evolving motor capabilities.