To overcome the issues of energy attenuation in fault features and modal overlap that traditional modal decomposition methods often encounter in high-noise environments, sparse reconstruction theory has been introduced into the field of bearing fault diagnosis. These methods achieve precise approximation of weak fault impacts by identifying the atomic combinations that best match the signal structure within an overcomplete dictionary, thereby mechanistically ensuring the integrity of feature energy. The Bayesian orthogonal matching pursuit (BOMP) algorithm demonstrates advantages in reconstructing and identifying weak features. However, existing methods typically eliminate atoms with contributions below a preset threshold during iteration, discarding the faint yet crucial fault feature energy contained within these atoms. This compromises signal reconstruction accuracy and noise reduction effectiveness. To address this issue, this paper proposes an improved Bayesian orthogonal matching pursuit (IBOMP) model. Its core enhancement lies in optimizing the decision threshold strategy for support set atom selection, aiming to maximize the retention of faint fault feature energy contained within low-contribution atoms. Comparative experiments against classical sparse reconstruction algorithms—including orthogonal matching pursuit (OMP) and Bayesian orthogonal matching pursuit (BOMP)—demonstrate that the proposed IBOMP method more effectively suppresses noise interference and enhances fault feature signals. Validation using the bearing fault dataset from the intelligent diagnosis and expert system laboratory at Nanjing university of aeronautics and astronautics demonstrates that compared to OMP and BOMP algorithms, the proposed IBOMP significantly enhances the frequency energy contribution of bearing rolling element and outer ring fault features—by 22.71%, 22.73%, 46.22%, and 46.52%, respectively.