This study proposes an embedded modeling methodology for identifying the crack-induced local stiffness reduction in a shaft from its horizontal and vertical vibrations. An embedded model integrating a dynamic model of the rotor system, and the unknown local stiffness reduction in the form of a universal function approximator i.e., neural network, is established. As a FEM model, it can describe rotors of complex geometry. A solution method is then established to identify the local stiffness reduction of the shaft due to a crack. Subsequently, a method is then used to find the location and size of the crack along the shaft. Simulated studies were conducted to demonstrate that the crack induced stiffness reduction of a Jeffcott rotor system can be identified, and the location, size and shape of the crack can be estimated by the proposed method with high level of accuracy.