A prediction model for upper limb motion direction based on the electromyography force mapping matrix was proposed to map the relationship between surface-electromyography (sEMG) signals and the direction of upper limb motion. Electromyography signals of the subjects were collected through experiments, and the accuracy of the force direction recognition algorithm for the upper limb in different postures was analyzed. Electromyography signals of nine superficial muscles were used as input, with wavelet filtering and root mean square（RMS）applied to process the signals. Additionally, the sEMG-force direction mapping matrix (SFMM) and the end motion direction mapping matrix (EDMM) were constructed. The model was trained by pseudo-inverse method and back propagation neural network (BPNN). The prediction performance of directly using the original data, the processed data, and the data combined with the posture transformation matrix in single posture and mixed posture was compared through experiments. The research results show that the method combined with the posture transformation matrix shows high accuracy in various postures, which can effectively reduce the influence of upper limb posture changes on the prediction results. This study provides a theoretical basis for the force direction prediction of exoskeleton devices based on sEMG.