Abstract: Structural health monitoring is gradually applied to the operation, to the management, to the maintenance, and to the safety assessment of structures, but there is a lot of noise and uncertainty in the massive monitoring data, which makes it difficult to achieve the accurate structural performance assessment. In this study, structural performance assessment enabled by Sparse Bayesian Learning (SBL) and by the improved D-S evidence theory for multi-data fusion is constructed. The spectral analysis was carried out for the measured acceleration data and strain data. The multi-data feature-level fusion is performed by using the matrix linear transformation technique, which allows the construction of the SBL regression model and the quantification of its residuals with the help of Bayesian hypothesis testing. Based on the measured acceleration and strain data, structural performance assessment is investigated by introducing weighting factors and by assigning different degrees of reliability between the evidences to further carry out the structural performance degradation localization analysis. Taking the measured monitoring data of Tsing Ma Bridge (TMB) as an example, the feasibility and effectiveness of structural assessment performance is verified upon Sparse Bayesian Learning and upon the improved D-S evidence theory for multi-data fusion. Study results show that: the method considers both acceleration and strain multi-data, and the fusion of the Bayesian factor shows an increasing trend over time, which can accurately identify the degree and location of the deterioration of structural performance, obtaining more comprehensive structural performance assessment information.