Resistance-Based Features Boost Machine Learning Fault Detection in Power Grids

Electrical faults in power systems can cascade into costly outages if they are caught too late, and a new study reports that a change in the underlying signal may make early detection more reliable. Researchers propose a hybrid framework that converts raw voltage and current measurements into phase-wise resistance values, then feeds statistically derived features from those resistance signals into machine learning classifiers for fault classification. The approach addresses a known weakness of conventional diagnosis techniques that rely directly on voltage and current waveforms.

Such raw measurements can be sensitive to measurement noise and to variations in operating conditions, which may degrade classification performance. From the phase-wise resistance signals, the team extracted five statistical features: the mean, maximum, minimum, standard deviation, and range. Six machine learning algorithms were evaluated on these features: Random Forest, Extreme Gradient Boosting (XGBoost), CatBoost, Support Vector Machine, Logistic Regression, and K-Nearest Neighbors.

The models were trained and tested using an 80:20 split, a common protocol in which eighty percent of the data trains the model and the remaining twenty percent measures its performance on unseen cases. Random Forest achieved the highest accuracy of the six classifiers. It reached 97.75 percent on the Electrical Fault Detection and Classification dataset hosted on Kaggle, and 95.82 percent on the Hybrid Electrical Fault Detection Dataset, which combines real and synthetic data and is available through Mendeley Data.

Testing on two independent datasets provides some evidence that the feature representation generalizes beyond a single data source. The consistency of the results across both datasets is notable because the two collections differ in composition. A feature set that supports high accuracy on both suggests the resistance-based descriptors are not narrowly tuned to one measurement environment.

The authors position the framework as a step toward predictive maintenance and intelligent fault monitoring. The study, published as an open-access article in Scientific Reports, is being shared early as a peer-reviewed, accepted version that remains subject to further edits before the final Version of Record. Its reported accuracies come from the two evaluated datasets under the specified training split, and the authors describe the framework as demonstrating potential for predictive maintenance rather than reporting field deployment results.

Subject of Research: Machine learning-based electrical fault classification using resistance-derived statistical features

Article Title: A hybrid approach for early fault prediction based on resistance behavior and machine learning

Article References: Charjan, O., Gajbhiye, K., Warhade, J., Wankhade, S., Mallick, S. R., & Wankhade, K. K. (2026). A hybrid approach for early fault prediction based on resistance behavior and machine learning. Scientific Reports. https://doi.org/10.1038/s41598-026-76547-0

Image Credits: AI Generated

DOI: 10.1038/s41598-026-76547-0

Keywords: fault detection, power systems, machine learning, Random Forest, resistance features, predictive maintenance, XGBoost, CatBoost, Support Vector Machine, electrical fault classification, hybrid, approach