Continuous prediction of knee joint angle in lower limbs based on sEMG: a method combining an improved ZOA optimizer and attention-enhanced GRU

Abstract

Exoskeleton robots have been increasingly applied in mountaineering, rescue, and military scenarios to alleviate physical burden and enhance mobility. This study proposes a novel approach for continuous knee joint angle prediction based on surface electromyography (sEMG), integrating an Improved Zebra Optimization Algorithm (IZOA) with an attention-enhanced Gated Recurrent Unit (GRU) network. The IZOA leverages Tent and Logistic chaotic mappings for improved population diversity and convergence, along with a memory-based strategy to enhance global search capabilities. Experimental evaluations across three motion tasks—level walking, stair ascent, and stair descent—demonstrated that the proposed method achieved a minimum root mean square error (RMSE) of 1.31°, with over 50% reduction in feature dimensionality, significantly outperforming Genetic Algorithm (GA), Zebra Optimization Algorithm (ZOA), Liver Cancer Algorithm (LCA), and Pied Kingfisher Optimizer (PKO). In addition, normalization based on maximal voluntary contraction (MVC) improved model robustness across subjects. The attention-based GRU further enhanced dynamic feature extraction, leading to an average RMSE reduction of 27.2% compared to baseline GRU and Long Short-Term Memory (LSTM) models. These results confirm the effectiveness of the proposed method in achieving accurate, stable, and continuous sEMG-driven knee joint angle prediction, offering strong potential for intelligent control in wearable exoskeleton systems.