Detection of broken rotor bar (BRB) faults in metro traction motors is challenged by two coupled difficulties: (i) under low-speed and light-load conditions, the characteristic sideband at twice slip frequency (2sf) is weak and masked by the fundamental; and (ii) under compound-fault scenarios, faults in other transmission components induce torque ripples that produce 2sf modulation in the motor current that is easily confused with genuine BRB. We propose a multi-dimensional current feature ensemble that combines demodulation of a squared-current instantaneous-power proxy with multi-slip 2sf amplitudes, wavelet subband energies, Teager energy, time-domain statistics, and a negative-sequence ratio, classified by an ExtraTrees ensemble. On the PHM-Beijing 2024 final-stage dataset (294 labeled samples, 42 BRB), the method achieves AUC 0.999, overall accuracy 99.3%, recall 97.6%, and a false-alarm rate of 0.4%; a lower operating threshold yields 100% recall at 98.3% overall accuracy. All figures are a post-hoc evaluation on the test ground-truth labels, which were unavailable during the competition; repeated nested cross-validation, with the decision threshold tuned strictly inside the training folds, retains 98.3% ± 0.4% accuracy at 98.8% ± 1.6% recall over ten repeats; a leave-one-speed-out evaluation attains AUC 1.00 when transferring between the 40 and 60 Hz conditions, while transfer to 20 Hz is limited by the sub-resolution 2sf line.

Broken rotor bar, motor current signature analysis, instantaneous power, Teager energy operator, wavelet decomposition, negative-sequence component, ExtraTrees, compound fault