Prediction of cast iron guide roller wear from multivariate data analysis using supervised machine learning: A case study with production process of construction steel bars

Most industries aim to replace traditional maintenance methods with more sophisticated, data-driven approaches. By incorporating the Industrial Internet of Things (IIoT), maintenance processes become even more efficient, providing predictive insights that help prevent downtime, streamline operations, extend asset lifespan, and optimize maintenance scheduling. Moving from traditional preventive maintenance to a more advanced, data-driven predictive maintenance (PdM) approach, this study investigates the application of machine learning (ML) to predict different sources of wear of cast iron guide rollers used in steel bar production. Leveraging historical data on guide roller replacements from the beginning of 2021 until the end of 2023, several operational parameters were analyzed, including vibration, overload, speed changes, product diameter and others. The problem was treated with two approaches. A linear approach taking the weight of the billets transformed into the final product as an indication of the remaining useful life (RUL) of the guide rollers before a critical level of wear occurs, and a discrete approach by subdividing the weight produced into five categories. For this reason, various ML algorithms, including linear regression (LR), support vector machines (SVM), random forest (RF), logistic regression (LogR), naive bayes (NB), decision tree (DT), and k-nearest neighbor (kNN), were applied to this dataset. The study identifies the primary causes of guide roller wear such as vibration, overload, and speed changes which lead to unplanned downtime, reduced product quality, and increased maintenance costs in industrial settings. It proposes predictive and condition-based solutions aligned with Maintenance 4.0 principles. A thorough comparison of these models was conducted to determine the most suitable algorithm for predicting roller degradation. The models were assessed based on the Mean Squared Error (MSE) and especially on the Root Mean Squared Error (RMSE) for the linear method. While for the multiclass classification method, performances were evaluated using accuracy, precision, F1 score, and recall. Model evaluation revealed the following finding: selecting relevant features from the dataset significantly improved performance for both the linear and discrete methods. The linear method achieved a prediction accuracy of approximately 94 %, while for the multiclass classification method, accuracy, precision, and F1 score reached 99.9 %.