Sparse filtering and its variants have been used in the field of weak feature extraction of bearing. However, the fault features extracted by the current sparse filtering methods are still subject to contamination of strong noises. Therefore, an enhanced sparse filtering fusion method is proposed for bearing fault diagnosis in this study. Specifically, the proposed method is conducted through the following steps. First, extract the fault features under background noise by an enhanced sparse filtering. Second, Gini index is used to select the sparser fault features in the extracted fault features for constructing a multi-dimensional enhanced fault feature fusion source. Third, obtain the intrinsic manifolds of the multi-dimensional enhanced fault feature fusion source by the manifold learning algorithm. Finally, the intrinsic manifolds are weighted to recover the fault-related transients. Analysis and comparison results of the experiment data from defective bearings indicates that the proposed method shows a prominent superiority in bearing fault diagnosis.