It is well known that rotating elements are subject to failure under fatigue circumstances, caused by crack propagation in high-cycle stress. Therefore, the operating monitoring of critical systems is essential to guarantee society’s well-being since the procedure can avoid unexpected downtimes and possible accidents in factories, vehicles, and power plants. Data-driven models emerge as prominent research lines for detecting faults and quantifying parameters in rotor machines. The popularization of these techniques could be directly linked to the massive capability of artificial intelligence algorithms to handle problems with multi-dimensional characteristics. However, to fully employ these techniques in critical systems, it is necessary to understand the prediction mechanism of the data-driven models produced during training. In this work, a signal and a model-based diagnostics method developed with Bayesian neural networks are applied and compared to determine the unbalance level, the size, and the crack’s location in a shaft. The methodology used simulated data from a rotor supported by two journal bearings. The breathing rotor crack model and the assumption of short bearings were used to generate synthetic data and train the models.