Artificial neural networks for predictive maintenance: fault prediction in belt drive systems

This study addresses the challenge of predictive maintenance in belt drive systems focusing on filling a specific gap by establishing a direct functional link between complex Environmental and Operational Conditions (EOCs) and system health indicators. A Feed-Forward Backpropagation Neural Network () model was developed to accurately predict four statistical vibration features: Variance (), Mean Absolute Deviation (), Energy (), and Zero-Crossing Rate (), with and being the most robust indicators. By employing temperature, rotational speed, and aging conditions as primary inputs, the framework is specifically tailored to accommodate the high variability associated with EOCs, a known challenge in diagnostics. The optimal architecture was systematically determined using the Intelligent Problem Solver (IPS) and validated through the Training-Validation-Testing (T-V-T) methodology to ensure superior generalization. Evaluation demonstrated a high degree of accuracy and robustness across all experimental conditions, with models for and achieving high correlation coefficients of approximately 0.965 and 0.959, respectively. This accurate quantitative prediction provides a validated and highly robust tool for early fault prognosis, enabling the dynamic definition of maintenance thresholds and facilitating proactive, timely interventions in industrial predictive maintenance applications where such comprehensive quantitative models for belt drives are currently limited.