Remote controlling using surface electromyography (sEMG) plays a more and more important role in a human-robot interface, such as controlling prosthesis devices, and exoskeleton. Different gestures are controlled by the cooperation of muscle groups, and sEMG represent the energy of the activated muscle fibers. With the limit of the low performance of wearable device, this article proposed a remote hand gesture recognized system based on deep learning framework of multi-attention mechanism convolutional neural network using sEMG energy to decoding hand gestures with remote server host. In the first part, an adaptive channel weighted method is proposed on multi-channel data of sEMG for enhancing the related feature map of sEMG and reducing the feature map low contribution of sEMG. The second part is improving the shortcuts by adding adaptively weighted instead of a simple short concatenation of feature maps. A novel multi-attention deep learning framework with multi-view (MMDL) for hand gestures recognition is proposed in our study, using sEMG. We verify the MMDL framework on myo dataset, myoUp dataset, and ninapro DB5, with the average accuracy 99.27%, 97.86%, and 97.0%, which is improved by 0.46%, 18.88%, 7% compared with prior works. In addition, the framework can classify seven hand gestures with 99.92% accuracy on ours datasets.