Machine Learning-Based Fault Prediction for Polyurethane Conveyor Belts in Pharmaceutical Applications

Abstract

Continuous transport systems, especially conveyor belts, play a crucial role in modern industries including pharmaceuticals, where operational reliability, hygiene, and minimal downtime are paramount. This study focuses on the experimental modeling and predictive analysis of polyurethane (PU) conveyor belts subjected to varying operational conditions. A structured Design of Experiments (DOE) was developed using three key input parameters—Load (kg), Drop Height (m), and Motor Current (A)—each tested at three discrete levels. Output responses included Vibration RMS (mm/s), Current Deviation (A), and Impact Force (kN), which serve as fault indicators. A dataset comprising 30 tests was generated based on deterministic equations to capture physical behavior realistically. Random Forest and Decision Tree regressors were trained and evaluated, achieving an R² score of 0.9995 and error metrics (MAE, MSE, RMSE) well below 1%, confirming the high accuracy of the predictive models. Correlation analysis and permutation-based feature importance revealed Load as the most influential factor, followed by Motor Current, while Drop Height had minimal impact on system responses. Radar plots and scatter graphs further validated the model’s prediction fidelity. The study not only demonstrates a reliable DOE framework for conveyor fault modeling but also establishes a scalable pipeline for real-time machine learning-based condition monitoring. These findings are particularly applicable to pharmaceutical industries where predictive maintenance and operational stability are critical.