Thermal models are often used in motor overload protection relays to predict rotor temperature. Under impaired cooling conditions, the parameters in the thermal models need to be tuned accordingly to reflect changes in the motor cooling capability.The tuning process requires accurate rotor temperature as a reference signal. This reference signal is estimated from the rotor resistance based on the equivalent circuit of the induction machine; however, it is often corrupted by noise due to the asymmetry in the power supply. A detailed analysis of the trajectories of the voltage and current space vectors in the synchronously rotating d-q reference frame is presented in this paper. The axis misalignment due to the negative sequence space vectors is identified as the major source of noise. Based on this analysis, a robust and efficient rotor temperature estimator is proposed in this paper. A digital positive sequence waveform synthesizer is designed to help align the d-axis with the positive sequence current space vector. Then, the induction machine positive sequence equivalent circuit is used to give a reliable estimate of the rotor resistance, which in turn yields the rotor temperature. This estimated temperature is subsequently used to tune the parameters of the rotor thermal model online. The proposed rotor temperature estimator is validated by experiments. The estimated rotor temperature is demonstrated to be sufficiently accurate, and is therefore appropriate to be used as a reference signal for the online tuning of the rotor thermal model parameters. The thermal model built from this reference signal is able to track the rotor temperature accurately under impaired cooling conditions.