Recovering high quality surfaces from noisy meshes is a fundamental problem in geometry processing. The main challenge is to robustly and efficiently handle different kinds of noise including impulsive noise, Gaussian noise while effectively preserving geometric features. In the paper, we first propose a normal filtering model with a novel aggregated fidelity term. The using of the aggregated fidelity term makes the proposed model robust against outliers and comparatively large noise. Then, a patch-aware normal filtering method is presented also based on the weighted least squares framework, which can preserve sharp features. Finally, a folding-free vertex updating scheme is employed to reconstruct the mesh according to the filtered face normals. Intensive experiments on a variety of surfaces demonstrate the superiority of our denoising method visually and quantitatively.