Surface electromyographic(sEMG) signals are considered to be the most suitable physiological feedback signal for acquiring human movement intention. Human-computer interaction technology based on sEMG signals is widely used in intelligent prosthesis control, sign language recognition, skeletal rehabilitation and other fields. Aiming at the problems of poor model stability and low precision in continuous hand motion recognition based on sEMG signals, a new deep learning method (BiLSTM-Att) based on LSTM and multi attention mechanism was proposed. Firstly, the RMS features of sEMG signals of different channels were extracted, and then the mapping model between motion and sEMG signals was constructed, and combined with the time-space information of signals, the historical information of depth features was mined. Finally, different feature subspaces were input to more stably and accurately predict the metacarpophalangeal joint angle in real time. BiLSTM-Att was evaluated on 6 finger movements of 15 subjects in the NinaPro DB2 dataset and compared with CNN-BiLSTM network, LSTM and TCN in the same operating environment. The experimental results show that the average CC, RMSE and NRMSE (0.85±0.03; 9.29±1.88; 0.10±0.01) was significantly better than that of CNN-BiLSTM(0.81±0.05, p<0.001; 10.12±2.09, p<0.001; 0.11±0.01, p<0.001), LSTM(0.75±0.06, p<0.001; 11.60±2.71, p<0.001; 0.13±0.02, p<0.001) and TCN(0.71±0.06, p<0.001; 12.40±2.60, p<0.001; 0.14±0.02, p<0.001), which indicates that BiLSTM-Att of the paper can estimate finger continuous motion more stably and accurately.