This paper presents a novel Bayesian Kriging model for quality design which tackles both variable uncertainty and model structure uncertainty in the metamodeling process, thus providing a robust foundation for quality improvement. Within the Bayesian hierarchical framework, significant variables in the global trend model of Kriging are effectively identified, and the validity of the candidate models is rigorously assessed through statistical tests. Initially, factor effect principles are integrated into the prior distributions of parameters to clarify their relationships, thus significantly reducing the dimensionality of the candidate space. Subsequently, Markov Chain Monte Carlo simulations are employed to estimate the posterior probabilities of the models, identifying Kriging models with sparse global trend structures. The validity of candidate models is then analyzed through multiple hypothesis testing, with corrections applied to the underestimated prediction variance. The final model is selected based on a comprehensive assessment of both its validity and generalization capability. The simulation results indicate that the proposed method performs satisfactorily across different sample sizes and significance levels. Additionally, the results of the case studies under two industrial scenarios demonstrate that the proposed method effectively identifies significant variables under both differential and non-differential posterior probability conditions.