Feature engineering for fault detection and diagnosis in Power Transmission Lines using a tree-based approach

The Power Transmission Line (PTL) is a pivotal infrastructure in the efficient distribution of electrical power, connecting the source generating stations to end-point consumers. This intricate system utilizes various components of diverse sizes and functionalities to transmit power. However, these components are susceptible to failures due to harsh environmental conditions and aging phenomena, posing a risk to the secure transmission and availability of electric power. In this study, data were gathered through the simulation of power transmission lines using MATLAB Simulink. The collected data were meticulously analyzed to discern the relationship between fault occurrences and input parameters. Subsequently, a preprocessing phase involved the introduction of new features as part of feature engineering. Six machine learning classifiers were employed in a bi-phased system: the primary objective was to detect faulty samples within the data, then diagnose these faults and distinguish their types. The results demonstrated the robust performance of LightGBM in detecting faulty samples, achieving an accuracy that exceeds 99.8%. In the second phase, the ExtraTrees classifier dominated, exhibiting a 99.8% accuracy and 100% classification precision in diagnosing four of the five fault types studied. Through machine learning explainability, it was revealed that the added features can significantly enhance prediction performance, particularly in the phase of fault diagnosis. The training time did not exceed 0.5 s for fault detection and 0.3 s for fault diagnosis. This swift convergence, coupled with the minimal training times, underscores the efficiency of the proposed feature engineering process and the selected classifiers. • Proposed a bi-phase tree-based framework for fault detection and diagnosis. • Conducted exhaustive data analysis to extract insights and engineer a novel feature. • Introduced transparency and interpretability using explainable machine learning.