Análise e aperfeiçoamento de modelos inteligentes para detecção de lâmpadas de iluminação pública

Abstract

A few years ago, a change in legislation transferred the responsibility for managing and maintaining the public lighting network from electric companies (now responsible only for billing the energy used) to municipal authorities. This change led to several misinformation problems, in which the electric company is often not notified about changes in the public lighting network. To avoid commercial losses, companies started to send manual conference teams, an expensive, time-consuming and unreliable process. In this sense, this work aims to improve the study of intelligent detection of street

lighting lamps, through the optimization of the models proposed by Soares et al. (2015), able to efficiently classify the type and power of street lighting point lamps as an alternative solution to this problem. The proposal is to use algorithms of different levels of complexity (from both traditional approach and deep learning), along with more complex techniques of validation, feature selection, data transformation and hyperparameter optimization. The results show that models with more complex algorithms (support vector machine, XGBoost, random forest and multilayer perceptron) manage to reach a final average accuracy of 80-86%, which when compared using Student's t tests did not show evidence of significant difference at the 5% level.