ResStarNet: A frequency-normalized and physics-guided framework for motor fault diagnosis under varying operating-frequency and load conditions
Measurement: Journal of the International Measurement Confederation, cilt.292, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 292
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.measurement.2026.123234
- Dergi Adı: Measurement: Journal of the International Measurement Confederation
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
- Anahtar Kelimeler: Continuous wavelet transform, Cross-condition generalization, Frequency normalization, Induction motor fault diagnosis, Lightweight residual CNN, Physics-guided learning
- Marmara Üniversitesi Adresli: Evet
Özet
This study presents ResStarNet, a lightweight residual CNN-based framework enhanced by frequency-normalized time–frequency representations and physics-guided regularization to improve motor fault diagnosis across varying speed and load conditions. ResStarNet integrates residual learning with a StarGate modulation mechanism to support transferable feature extraction while maintaining low computational complexity. To reduce the spectral mismatch that becomes dominant under large operating-frequency changes, raw motor-current signals are transformed into time–frequency representations using the continuous wavelet transform (CWT). The resulting scalograms are subsequently mapped into a normalized-frequency domain by expressing the frequency axis relative to the estimated fundamental frequency. This normalization establishes a consistent spectral reference frame and improves the alignment of dominant fault-related structures across different operating conditions. A class-specific physics-guided regularization strategy is further introduced in the normalized-frequency domain, in which candidate spectral intervals derived from analytical fault-frequency relations are refined using source-domain training data and used to construct soft spectral masks for regularizing intermediate feature representations. This guides the network toward physically consistent representations while maintaining tolerance to slip variation, spectral spreading, and estimation uncertainty. The proposed framework is evaluated on an experimental motor-current dataset containing healthy, rotor-fault, stator-fault, bearing-fault, and eccentricity-fault conditions under different operating conditions. The results show that ResStarNet supports transferable feature extraction under moderate and combined operating-condition changes. Frequency normalization provides the dominant improvement under large operating-frequency changes by compensating for speed-induced spectral displacement, whereas physics-guided regularization provides additional benefits under load-related cross-condition variability.