Abstract
<jats:p>Current implementation of predictive maintenance systems for industrial equipment, particularly belt conveyor systems, is hindered by technical complexity and the necessity to adapt solutions to specific technological processes. Existing methods often require profound knowledge in the field of vibration diagnostics or substantial computational resources. The relevance of this research is justified by the urgent need to develop accessible and efficient monitoring tools capable to detect faults (loop breakage, chain sagging) at early stages without production shutdown. The aim of this study is to develop and validate a predictive maintenance system for curved belt conveyors based on the concept of a digital twin and machine learning methods. Experimental investigations have been conducted on a laboratory test rig equipped with a comprehensive sensor array (accelerometers, microphones, current sensors). Data acquisition has been performed at high sampling frequency (51,2 kHz), followed by down sampling to 10 kHz using linear interpolation to optimize computational costs. For data analysis and condition classification (“normal” vs. “fault”), four machine learning algorithms have been implemented and compared. They are Random Forest, logistic regression, support vector machines (SVM), and decision trees. Model training and testing have been performed on a combined dataset with 80/20 train-test split ratio. Comparative analysis has demonstrated that the Random Forest algorithm exhibits superior performance with an AUC of 0,87 and F1-score of 0,88. It has been established that reducing the sampling frequency to 10 kHz represents an optimal trade-off, reducing the processing time for one minute of recorded data from 241 s to 56 s while preserving signal diagnostic value. The developed model successfully identifies various fault types, achieving high prediction accuracy (up to 98 % on selected test sets). The research results confirm the effectiveness of applying a digital twin approach combined with machine learning methods to diagnose electromechanical conveyor systems. The achieved accuracy and response time of the system allow us to recommend the proposed solution for implementation in real-world production processes. Application of the developed methodology enables transition from scheduled maintenance to condition-based maintenance, minimizing downtime and reducing operational costs without requiring specialized vibration diagnostics experts.</jats:p>