Analysis and Improvement of Machine Learning Models for Detecting Street Lighting Lamps

Abstract

The Brazilian public lighting network is maintained by city halls. To bill the energy provided to city halls, energy distribution companies should maintain an updated database of network poles, their lamp types, and wattages. However, it is common to encounter issues with misinformation, where the company is not notified about changes in the public lighting network by city halls and cannot update its database appropriately. To mitigate commercial losses, companies have resorted to sending teams for manual infrastructure checks, which is an expensive, time-consuming, and unreliable process. In this regard, this work aims to optimize the models proposed in the literature capable of accurately classifying the type and wattage of lamps on public lighting poles based on data collected from radiometric sensors and a professional camera. Data is processed using traditional machine learning and deep learning algorithms, along with more sophisticated validation techniques such as data transformation and hyperparameter optimization to achieve improved results. Based on this methodology, the results demonstrate that models employing more robust algorithms (Support Vector Machine, XGBoost, Random Forest, and Multilayer Perceptron) can attain a final average accuracy of 80-86%. This confirms the usefulness of this methodology as an alternative solution to address the issue of public lighting billing.