Deep Learning-Based Predictive Maintenance Framework Using CNN-LSTM Architecture for Industrial Rotating Machinery Fault Detection

Unplanned equipment failure in industrial manufacturing environments results in significant economic losses — estimated at USD 50 billion annually across global manufacturing sectors — arising from production downtime, emergency maintenance costs, and secondary equipment damage. Conventional time-based preventive maintenance schedules mitigate catastrophic failure risk but are operationally inefficient, frequently replacing components before their functional end-of-life and incurring unnecessary maintenance expenditure. Condition-Based Monitoring (CBM) through vibration, acoustic, and thermal sensor data offers a route to maintenance scheduling that is both reactive to actual machine health and predictive of imminent failure. The present study proposes and validates a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) architecture for multi-class fault detection in industrial rotating machinery — specifically centrifugal pumps, induction motors, and gearboxes — using time-series sensor data collected over 18 months at a precision engineering facility in Haryana, India. The CNN sub-network extracts spatial features from short-time Fourier transform (STFT) spectrograms of vibration signals, while the LSTM sub-network models temporal dependencies in multi-channel sensor streams including temperature, current draw, and acoustic emission. The proposed CNN-LSTM model achieves 95.6% classification accuracy, 94.8% precision, and an AUC of 0.981 across five fault classes (normal, bearing fault, shaft misalignment, cavitation, and lubrication deficiency) on a held-out test set of 4,400 labelled samples — outperforming standalone Support Vector Machine (87.3%), Random Forest (91.2%), and LSTM (93.8%) baselines. Real-time deployment on an NVIDIA Jetson AGX Xavier edge inference platform achieves end-to-end latency of 23 ms, suitable for safety-critical online monitoring applications. The framework reduces false positive maintenance alerts by 38% relative to threshold-based alarm systems and projects a 27% reduction in annual maintenance cost over a three-year evaluation horizon.