Dynamic displacement is an important physical quantity in the fields of seismic engineering, military weapon design, and structural health monitoring. In the actual test process, the acceleration can usually be directly measured. Due to the uncertain test conditions such as the environment, the acceleration signal is unavoidable contains low-frequency and high-frequency noise, which causes a significant drift in velocity and displacement during the acceleration integration process. Therefore, great scientific research and practical engineering significance to obtain a reasonable and scientific acceleration-displacement relationship. Based on the theoretical framework of Bayesian inference, a machine learning dynamic displacement identification method is constructed. The results show that, the displacement response obtained by inversion for different noise conditions (white noise, artificial noise) is basically consistent with the analytical displacement; the displacements of inversion of acceleration sensor signals with different performances are compared by using a large shaking table test data, and their uncertainty are analyzed. The results show that this method has certain advantages in the characterization of the acceleration-displacement relationship, and can achieve the displacement solution without relying on the processing of the acceleration signal, thereby avoiding the displacement integral distortion caused by the accumulated noise error.