Rehabilitation training can effectively help the elderly recover their self‐care state and enhance physical fitness. As surface electromyography analysis is effective to recognize motion intention, researchers use it to develop prosthetic limb operations. In this article, the rehabilitation training bed is designed by combining the rehabilitation training with the motion prediction based on the surface myoelectric signal, which can recognize the tilt of the upper body in different directions and provide corresponding assistance to the elderly. After collecting EMG signal, the effective signal was dimensional reduction, mapped by linear discriminant analysis. To train and recognize the EMG‐motion mapping relationship, we used a recurrent neural network called nonlinear autoregressive with exogenous input model and used a 360° tilt prediction experiment on the upper body. Results showed that the root mean squared error and the error autocorrelation coefficient were relatively low, and the tilt degree of the experimenter was highly matched.