This paper presents a localized mechanical fault detection method using spectral identification techniques. Two algorithms are evaluated, the Levenberg-Marquardt algorithm which is an iterative Nonlinear Programming algorithm and the ESPRIT algorithm based on parameter estimation via Rotational Invariance Techniques. A reduced frequency model is used to provide a very accurate estimation of the fault and to significantly reduce the calculation time. A dedicated experimental setup based on a novel mechanical fault emulator and a 7.8kW PMSM drive is designed to validate the proposed approach. The experimental results show that the proposed technique is effective to detect single mechanical fault signatures by evaluating a simple online fault indicator and it proves also its ability to detect early mechanical faults. Finally, a critical comparison has been made in order to evaluate their performance in terms of calculation time, complexity, advantages and drawbacks.