The action recognition technology based on surface electromyography (sEMG)signals has broad application prospects in perfect integration of human-machine interaction

but its inherent instability and latency remain key difficulties in this field.To overcome these challenges

a model based on multi-stream convolutional attention with gated recurrent unit network (MS-CAGRU-Net)is proposed.Firstly

a multi-stream dilated convolutional layer integrated with attention mechanism is introduced to dynamically adjust the weights of different locations or channels

thereby capturing subtle differences in input data.Secondly

the features of each branch

are fused to capture effective information in the signals more comprehensively.Lastly

a gated recurrent unit(GRU)is utilized to further capture long-term dependencies in temporal sequences.Experimental results indicate that the MS-CAGRU-Net model achieves an accuracy and an

F

1

score of 97.89% and 97.88%

respectively

and outperforms other comparative models in stability and reliability

which effectively improves the accuracy of action recognition.