An FFT-LPS–MSDRNet-1D–DAN Framework for Three-Phase Stator Current-Based Cross-Speed Fault Diagnosis of Induction Motors

To address weak fault signatures, unavailable target-speed sample labels, and spectral distribution shifts between source- and target-speed domains in current-only cross-speed fault diagnosis of induction motors, this paper proposes an unsupervised domain adaptation method integrating FFT-LPS, MSDRNet-1D, and DAN. First, three-phase stator-current window samples are transformed into FFT log-power spectra to highlight sidebands around the fundamental frequency, harmonic structures, and local band disturbances. Next, MSDRNet-1D, a multiscale depthwise-separable residual network for one-dimensional current spectra, extracts local spectral peaks, neighboring-band disturbances, and cross-band correlations. Finally, a DAN mechanism based on multi-kernel maximum mean discrepancy aligns deep feature distributions under labeled source-speed and unlabeled target-speed conditions. Experiments on the self-built motor-current dataset and the IEEE DataPort public three-phase induction-motor broken-rotor-bar dataset show that the proposed framework achieves the best overall performance among the compared methods and remains effective under noise and load variations. The results indicate that compact spectral representation, progressive multiscale feature extraction, and statistical domain alignment are suitable for current-based motor fault diagnosis under variable operating conditions.