Rolling bearings are essential components in a wide range of rotating machines, these components are critical for susceptibility to progressive wear and crack propagation. In addition, it is desirable to reduce the interruption and cost of preventive maintenance. Thus, accurate prediction of the remaining useful life of rolling bearings is crucial to efficient and safe operation. Rolling bearings work under cyclic load due to the elements orbiting around the center. The contact stress is caused by the external load and the corresponding bearing reaction. These periodic compressive and shear stresses are modeled by the Hertzian theory, and after numerous stress cycles, the fatigue process initiates the crack and its consequent time-propagation, leading to pitting and subsequent spalling. A comparative study of rolling bearings crack propagation is proposed. Fatigue life, as a function of fault growth, is approached by classical Paris’s law, assuming an incipient or initial crack and a maximum acceptable crack length. At this point, the numerical bearing simulations taking account of consecutive fault lengths and corresponding life cycles are applied to train a Long Short-Term Memory (LSTM) network to develop a predictive model for rolling bearing failure. This network detects the gradual progression of cracks over time, learning and adjusting its predictions as the fault length increases. The model-based LSTM and the classical time-life predictions are compared, enabling effective predictive maintenance. The method presents a promising solution for machinery health monitoring, providing an advanced alternative for prognosis management in high-demand industrial environments.