Comparação de algoritmos de machine learning para predição de vibração em moinhos de rolos verticais aplicados à manutenção preditiva industrial

Abstract

Predictive maintenance based on vibration monitoring has become a fundamental strategy for optimizing operational costs and increasing the availability of critical equipment in the cement industry. Vertical roller mills exhibit characteristic vibrational patterns that can indicate the development of operational anomalies before critical failures occur. This work presents a comparative analysis of six machine learning algorithms for vibration prediction in vertical roller mills, using real industrial data collected over 82 days of continuous operation, totaling 23,538 samples after preprocessing. The evaluated models were: K-Neighbors, Linear Regression, XGBoost, CatBoost, Huber Regressor, and Random Forest. The methodology included careful selection of 15 predictor variables based on operational knowledge and statistical analysis, feature engineering with categorical variable encoding, temporal dataset splitting for realistic validation, and comparative evaluation using quantitative metrics (R², RMSE, MAE, MAPE) complemented by visual analysis of time series and residuals. The results revealed that KNeighbors showed superior performance with R² of 0.7679, RMSE of 0.8792 mm/s, and MAE of 0.66 mm/s, surpassing traditionally more robust ensemble models. Visual analysis identified a critical regression-to-the-mean phenomenon in five of the six models, where high vibration values are systematically underestimated, a limitation that compromises applicability in alarm systems for predictive maintenance. K-Neighbors was the only model capable of preserving the natural variability of vibration, an essential characteristic for reliable detection of operational anomalies. This work demonstrates that model selection for critical industrial applications cannot be based exclusively on aggregate metrics, requiring qualitative evaluation of predictive behavior across the entire operational range. The developed methodology represents a replicable process for predictive systems in critical rotating equipment, with potential for significant reduction of unplanned shutdowns and optimization of operational costs.