EUVIMEDOpdag EUVIMED
Open access · CC BY 4.0

Beyond Benchmark Accuracy in Transmission-Line Fault Classification: Blocked Evaluation, Physical Identifiability, and Robustness to Wireless Sensor-Node Loss

Ali Olamat, Wasan Alamro, Muneera Altayeb

Computers · 2026-11-15 · 15 (11) · 835

DOI: 10.3390/computers15110835 ↗

At a glance

Journal article in Computers, published online on 2026-11-15. Written by Ali Olamat and 2 co-authors. 48 references. The full text is freely available (open access, CC BY 4.0).

AbstractBibliografiske oplysningerCiteRelated studies

Abstract (original language)

Machine-learning classifiers for transmission-line faults often reach near-perfect accuracy on simulated benchmarks, usually under random splits of temporally ordered samples. This study tests whether this holds under deployment-oriented conditions on a widely used six-class benchmark (7861 samples). Random and blocked (guard-gapped) 10-fold cross-validation compared five classifiers and conventional phase-selection logic; feature groups were ablated; robustness to noise and wireless sensor-node loss and conformal-prediction coverage were quantified. With the three-phase classes merged, random splitting yielded 99.93% accuracy versus 96.2% under blocking. The two three-phase classes had near-identical phase currents and line-to-line voltages, differing only in residual current (0.70 A offset; inception transient): a simulation artifact removed by phasor filtering and lost at signal-to-noise ratios of 30 dB (clean training) or 20 dB (noisy training). Causal window features, not physics-informed ones, raised blocked six-class macro-F1 from 0.893 to 0.967. Untrained relay logic matched learned classifiers on clean data (five-class macro-F1 0.989 versus 0.979), whereas the learned classifier decided earlier within fault records and, with masking augmentation, better tolerated node outages (0.958 versus 0.867 for a phase-B outage). Split-conformal prediction under-covered blocked test data (88% at a 95% target). High benchmark accuracy alone shows no advantage over conventional protection; independent fault events remain untested.

Cite

APA

Olamat, A., Alamro, W., & Altayeb, M. (2026). Beyond Benchmark Accuracy in Transmission-Line Fault Classification: Blocked Evaluation, Physical Identifiability, and Robustness to Wireless Sensor-Node Loss. Computers, 15(11), 835. https://doi.org/10.3390/computers15110835

Vancouver

Olamat A, Alamro W, Altayeb M. Beyond Benchmark Accuracy in Transmission-Line Fault Classification: Blocked Evaluation, Physical Identifiability, and Robustness to Wireless Sensor-Node Loss. Computers. 2026;15(11):835. doi:10.3390/computers15110835

BibTeX
@article{Olamat2026,
  author = {Olamat, Ali and Alamro, Wasan and Altayeb, Muneera},
  title = {Beyond Benchmark Accuracy in Transmission-Line Fault Classification: Blocked Evaluation, Physical Identifiability, and Robustness to Wireless Sensor-Node Loss},
  journal = {Computers},
  year = {2026},
  volume = {15},
  number = {11},
  pages = {835},
  doi = {10.3390/computers15110835}
}
Start for free →