10 July 2025 in IEEE Sensors Journal
Exoskeleton robots serve as crucial tools for active rehabilitation training among patients with lower limb motor dysfunction. Accurately recognizing human motion intention poses a significant challenge for exoskeleton robots, which can be addressed by continuously estimating human joint angles. Nevertheless, most existing studies primarily focus on healthy subjects, neglecting the target audience of lower limb exoskeletons, namely, patients with lower limb hemiplegia. This paper proposed a general model utilizing a transfer learning (TL) strategy to predict the gait trajectory of hemiplegic patients. Firstly, a gait dataset is established featuring hemiplegic patients walking on level ground. The original sEMG signal sequence is then decomposed using the empirical mode decomposition (EMD) algorithm to mitigate noise and non-stationarity inherent in the data signals. Secondly, kernel principal component analysis (KPCA) is employed to extract key features from the sEMG signal sequence, eliminating redundant information from the original sequence. Thirdly, considering the impact of spatio-temporal correlation in gait data on prediction accuracy, a prediction model based on CNN-BiLSTM is devised. Finally, TL technology is applied between the healthy and affected sides of the subject to transfer knowledge learned from the source domain data to the target domain. This method enhanced the accuracy of gait trajectory prediction for hemiplegic patients, alleviated the training burden, and bolstered the model’s generalization. The findings indicated that this TL strategy holds significant application value for advancing the development of lower limb exoskeleton control systems.