This work focuses on a data-driven non-intrusive proper generalized decomposition(PGD) surrogate model for parametric flexible mechanisms with adaptive sampling strategy. Similar to tensor decomposition methods, the non-intrusive PGD expresses the high dimensional dataset using the production of low-rank modes. The discrete data modes are obtained through the training process with the greedy enrichment strategy and the alternative least square method. To predict the dynamic response of flexible mechanisms with parameters out of the trainning set, the Kriging interpolation is used to find the optimal curve of the corresponding parametric mode. The Kriging model proposes the information of uncertainty about the optimal curves according to the exsiting dataset, and the contribution of each PGD modes to approximating the dataset can be regarded as the improtance of corrsponding curves. Thus, combining with a cross-validation process, an aquisition function for adaptive sampling can be constructed using the knowledge of Kriging model and the importance of PGD modes. After the adaptive sampling process reaches its convergence, the construction of PGD surrogate model is done. The proposed PGD surrogate model is validated through numerical example, where a flexible beam modelled by absolute nodal coordinate formulation(ANCF) is driven by a rotary motor. The input of the control strategy will also be regarded as the input parameter. The result shows the PGD model with Kriging interpolation have good performance for the predicting of the dynamic response of parametric flexible mechanisms with adaptive sampling strategy.