Análise de confiabilidade utilizando distribuições de sobrevivência e algoritmos de machine learning: um estudo de caso sobre injetoras na indústria calçadista

Abstract

The reliability of injection molding machines in the footwear industry is crucial, with accurate failure prediction being a significant challenge for predictive maintenance. This study examines the comparison of survival analysis methods in this context. The objective is to compare the predictive accuracy of traditional (parametric) approaches and Machine Learning (ML) techniques in estimating failure probability, identifying the most suitable methodology. Survival analysis is applied to real maintenance data collected between January 2023 and December 2024. Weibull AFT, Lognormal AFT, Random Survival Forest (RSF), and Gradient Boosting Survival (GBS) models are implemented and optimized. Accuracy is assessed using metrics such as C-index and Brier Score on test data, and the importance of variables in ML models is analyzed. The results indicate the predictive superiority of ML, particularly GBS, over the parametric models tested, demonstrating greater flexibility in adapting to observed data patterns. It is concluded that ML is a promising alternative for optimizing failure prediction and maintenance planning in this industrial scenario.