Abstract: A scheme of continuous motion estimation based on surface electromyography (sEMG) and long-short-term memory (LSTM)network is proposed for the control of lower limb assisted exoskeleton. The mapping relationship between EMG and motion is trained and analyzed by LSTM. The number of principal components (dimensionality reduction) for principal component analysis (PCA) algorithm are obtained based on the error algorithm of singular value decomposition eigenvalue matrix. The continuous motion estimation of three lower limb joints in sagittal plane is realized， and the real-time performance of the continuous motion estimation is improved. In comparison of the training results of LSTM network with those of traditional networks such as support vector machine (SVM) and back propagation(BP) neural network， the superiority of LSTM network in continuous motion prediction of lower limbs is proved.

Key words: surface electromyography(sEMG), long-short-term memory(LSTM) network, principal component analysis(PCA) algorithm, continuous motion estimation, real-time performance