To improve the accuracy of force/torque loading and the feasibility of multi-degree-of-freedom combined loading during the calibration of six-axis force sensors, a dynamic calibration system based on impact-hammer excitation was developed by utilizing a six-degree-of-freedom independent loading static calibration apparatus. The first- and third-order modal frequencies of the sensor structure were obtained through modal analysis to evaluate its dynamic response under multi-frequency excitation, thereby avoiding resonance effects and improving dynamic calibration accuracy. To assess the performance of the proposed calibration system, six-dimensional force/torque data were collected under known loading conditions. Three modeling and fitting methods—traditional linear regression, particle swarm optimization-based backpropagation neural network (PSO-BP), and support vector regression (SVR)—were employed to establish the mapping relationship between the sensor outputs and the applied loads. The fitting performance of the three methods was compared across all six channels. The results indicate that the PSO-BP neural network achieved the highest fitting accuracy, with the mean absolute error (MAE) of 17.86%，which reduced by 8.20% compared to linear regression, and significantly more than the SVR method. The proposed calibration system demonstrates excellent static and dynamic performance, validating the superiority and practicality of the PSO-BP neural network in multi-dimensional force/torque sensor calibration.