Artificial Intelligence Techniques for Predictive Maintenance in Smart Manufacturing Systems

Abstract

Predictive maintenance (PdM) is a central capability of smart manufacturing because it converts continuously collected machine-condition data into early warnings and maintenance decisions before functional failure occurs. This methodology paper develops an artificial-intelligence-based PdM framework that combines sensor acquisition, data cleaning, health-feature construction, supervised learning and maintenance decision support. The framework evaluates logistic regression, support vector machine (SVM), random forest and multilayer perceptron (MLP) models for prediction of a near-failure maintenance horizon. To demonstrate the methodology without claiming unavailable industrial plant measurements, a controlled synthetic smart-manufacturing benchmark of 6,000 multivariate observations was generated from vibration, temperature, motor current, hydraulic pressure, acoustic level, operating load and runtime variables. A stratified 75:25 train-test design and imbalance-aware learning were used. In the illustrative evaluation, random forest achieved the highest accuracy (95.2%), F1-score (0.863) and ROC-AUC (0.929), while SVM showed slightly higher failure recall (0.850) than random forest (0.828). The results support the use of ensemble and nonlinear AI methods when interactions among machine-condition variables are important. The proposed procedure is reproducible, technology-neutral and suitable for later validation with real factory data.