Intelligent optimal preventive replacement maintenance policy for non-repairable systems

[Display omitted] •We experiment both analytical and intelligent models for optimal preventive maintenance policies.•A loglogistic distribution with scale and shape parameters was identified to guide the replacement models.•Feature engineering uncovered key features for optimizing machine parameters, enhancing the intelligence of the selected models.•Replacement model based on the failure distribution function outperformed the hazard-based ones; while the selected machine learning models (XGBoost, SVM, RF, MLP) yielded more precise accuracy.•The study encourages practical implementation and validation of proposed policies in diverse industrial settings. This paper developed optimal maintenance policies for industrial machines to ensure uninterrupted production. Analytical and intelligent replacement models were constructed using real-time data from a commercial photocopier operator’s expenditure logbook between 2012 and 2023. A goodness-of-fit test affirmed a log-logistic failure distribution with shape and scale parameters of 1.7233 and 763.9220, respectively. Probability functions from this distribution established replacement models, determining optimal preventive maintenance conditions. Through feature engineering additional features crucial for operational and cost optimization were abstracted, with 4,881 unique data points simulated from boundary conditions defined by the original or raw data. Analytical results showed that the failure distribution-based replacement model outperformed the hazard function-based model in minimum replacement time and unit maintenance cost over time, while key parameters such as availability, reliability, and failure occurrence probability remained constant at 96 %, 94 %, and 0.07 %, respectively. Machine learning models (Extreme Gradient Boosting: XGBoost, Support Vector Machine: SVM, and Multilayer Perceptron: MLP) demonstrated very high accuracy in availability (XGBoost = 99.49 %, SVM = 99.18 %, RF = 100 %, MLP = 99.80 %), reliability (XGBoost = 99.49 %, SVM = 99.18 %, RF = 100 %, MLP = 99.80 %), and maintenance cost (XGBoost = 100 %, SVM = 100 %, RF = 100 %, MLP = 100 %), exceeding results of the analytical models.