This paper focuses on induction motor monitoring based on stator current measurements. The diagnosis aims at identifying the mechanical faults related to either airgap eccentricity or load torque oscillation. The airgap eccentricity (respectively load torque oscillation) essentially results in an amplitude (respectively phase) modulation of the stator current. Classical spectral analysis allows for the detection but not for the discrimination of these modulations. Time-frequency representations, such as the spectrogram or the Wigner distribution, provide appropriate signatures for fault discrimination. This paper proposes to perform the decision task from the time-frequency representation using the surrogate data technique. In a deterministic context, the phase and amplitude modulations can be considered as non-stationarities since they correspond to time-variations of the signal spectral content. The detection of a modulation is expressed as a binary hypothesis test. The null hypothesis corresponds to a signal without modulation. Stationarized/unmodulated replicas of the observed (possibly modulated) signal are obtained by phase randomization of its Fourier transform. These so-called surrogates provide a reference for the null hypothesis. The observed signal is then compared to these surrogates using appropriate distances in the time-frequency domain. A one-class classifier may be used considering the surrogates as a learning set. This classifier detects outliers corresponding to the modulation and thus to the failures. Moreover, this technique provides the information concerning the predominant type of modulation. This diagnosis method will be tested on simulated and experimental signals.