Signal-Driven Fault Detection in Electrical Systems via Adaptive Machine Learning

Abstract

Accurate and timely detection of faults in power systems is critical for ensuring operational reliability and safety. This study proposes a machine learning based fault detection framework that utilizes statistical features extracted from voltage and current time series signals. Each sample consists of one current row and one voltage row, from which fourteen time domain features are computed, including mean, standard deviation, skewness, kurtosis, root mean square, energy, and peak to peak values. These features serve as inputs to six adaptive classifiers: Gaussian Naive Bayes, K-Nearest Neighbors, Support Vector Machine, Random Forest, AdaBoost, and XGBoost. Experimental results show that ensemble methods, particularly Random Forest and XGBoost, significantly outperform others, achieving up to 99% accuracy and F1 scores in detecting both normal and fault conditions. The findings highlight the effectiveness of ensemble learning in handling noisy and imbalanced electrical signal data, and demonstrate the feasibility of deploying such models for real time monitoring in smart grid environments.