This paper explores the design, online, of an electrical machine's healthy reference by means of statistical tools. The definition of a healthy reference enables the computation of normalized fault indicators whose value is independent of the system's characteristics. This is a great advantage when diagnosing a broad range of systems with different power, coupling, inertia, load, etc. In this paper, an original method called spectral kurtosis with reference is presented to design a system's healthy reference. Its principle is first explained on a synthetic signal. This approach is then evaluated for mechanical unbalance detection in an induction machine using the stator current instantaneous frequency. The normalized behavior of the proposed indicator is then confirmed for different operating conditions, and its robustness with respect to load variations is demonstrated. Finally, the advantages of using a statistical indicator based on a healthy reference compared with a raw fault signature are discussed.