Electrical Current Signature-Based Machine Learning Models for Streetlight Fault Prediction in Smart City Infrastructure

Abstract

Urban street lighting systems are critical infrastructures, yet their maintenance is predominantly reactive, inefficient, and costly. The paper introduces a proactive, data-driven framework using machine learning (ML) to predict incipient streetlight faults by analyzing electrical current signatures. This methodology, which leverages features like current fluctuations and voltage variations as early indicators of failure, is significantly underexplored in smart lighting infrastructure. We utilize a public Kaggle dataset encompassing operational parameters and annotated fault types. To overcome the absence of high-resolution current data, we augmented the dataset with simulated current signature features (e.g., mean current, variance, power factor) engineered to reflect typical electrical behaviors under various fault conditions. Recognizing that fault data is inherently imbalanced, we applied the Synthetic Minority Over-sampling Technique (SMOTE) to the training set to prevent model bias and improve the detection of rare fault classes. A comprehensive comparative analysis of eight supervised ML models; Decision Tree, Random Forest, Logistic Regression, SVM, K-Nearest Neighbors, XGBoost, Naive Bayes and AdaBoost, was then conducted to identify the most robust approach. The results, validated via 5-fold cross-validation, show that ensemble methods are superior in capturing the data's complex, non-linear patterns. Random Forest achieved the highest classification accuracy (84%), with strong supporting metrics (Precision: 0.87, Recall: 0.85). Notably, AdaBoost demonstrated the best discriminative ability (AUC = 0.79), highlighting a practical trade-off between overall accuracy and threshold-tuning flexibility. A feature importance analysis confirmed that electrical signature features were among the most influential predictors. This research validates electrical current signature analysis as a viable methodology for fault prediction, providing a framework to shift maintenance from a reactive to a proactive strategy. We acknowledge the limitation of using simulated data and position this work as a foundational proof-of-concept. The study provides an empirically validated baseline for future intelligent maintenance systems, with the clear next step being validation on field-collected, non-simulated sensor data.