In continuous motion prediction using surface electromyographic (sEMG) signals, performance is strongly affected by combined noise such as electrode position. Most existing research has focused on neural network layers structure, making generalisation performance highly dependent on hyperparameters. To effectively address this issue, improvements are made from an intra-layer perspective. In this paper, based on the theoretical analysis, a neural network module with noise suppression capability called Luen-Mamba is presented. This module incorporates the Luenberger observer and adds output information to assist the network in reconstructing the system state. Comparison experiments show that Luen-Mamba predicts results more accurately than other neural networks. Finally, a comparison of the upper limb continuous motion prediction results shows that Luen-Mamba is effective in suppressing integrated noise.