Industrial rotating machinery such as motors, pumps, gearboxes and bearings account for a significant share of unplanned downtime losses in manufacturing and process industries globally, with annual costs exceeding USD 647 billion. Conventional rule-based condition monitoring approaches dependent on fixed vibration amplitude thresholds fail to capture the complex time-frequency interactions that characterise incipient fault signatures. This study proposes and evaluates a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) architecture trained on raw time-domain vibration signals from a purpose-built experimental rotating machinery rig equipped with four accelerometers sampling at 25.6 kHz. The proposed model simultaneously learns spatial feature hierarchies via convolutional layers and temporal dependencies via LSTM units, eliminating the need for hand-crafted features. Five fault classes—bearing inner race defect, rotor imbalance, gear tooth wear, stator winding fault, and normal operation—are classified under variable speed (600–1800 RPM) and variable load (0–75% full load) conditions. The CNN-LSTM model achieves overall classification accuracy of 97.3%, outperforming standalone Support Vector Machine (84.1%), Random Forest (89.4%), and Artificial Neural Network (92.6%) baselines. The proposed architecture demonstrates robustness across operating speed variations and maintains above 96% accuracy at signal-to-noise ratios as low as 5 dB, confirming its suitability for deployment in industrially realistic noisy environments.
Rajesh Kumar Sharma, Meenakshi Devi, Suresh Chandra (Sun,) studied this question.