As an indispensable and crucial component of advanced equipment, the running status of the bearings directly affects the operational safety of the entire system. By monitoring the vibration signals on the surface of the equipment, their abnormality can be effectively detected for real-time fault diagnosis. The recovery of fault signature in the early stage is therefore significant for industrial applications. This paper employs the improved sparse low-rank model (ISLR) for the reconstruction of the vibration signals. It proceeds from the property that, in the time–frequency domain, the fault signals exhibits pronounced sparse and low-rank structure. The selection of hyperparameters is conducted through a new procedure by combining the grid search with the cross-validation, guided by the envelope entropy of the reconstructed signal toward the symptomatic structure. The performance of the optimization method is validated by a synthetic signal, and its superiority is further demonstrated by a comparison with that using default parameter settings. Results show that the proposed hyperparameter optimization is distinctly improved the performance of the ISLR model in terms of noise-reducing ability.