CNN-LSTM Hybrid Deep Learning Architecture for Predictive Fault Diagnosis in Rotating Machinery: Multi-Sensor Vibration Analysis and Real-Time Deployment in Industrial Environments

Abstract

Rotating machinery faults including bearing defects, gear damage, shaft imbalance, and misalignment are the primary cause of unplanned downtime in manufacturing, mining, and power generation facilities, collectively responsible for equipment downtime costs estimated at USD 50 billion annually in industrialised economies. Predictive maintenance systems based on vibration signal analysis and machine learning have demonstrated strong fault detection capability in controlled laboratory settings, but real- world industrial deployment presents additional challenges of multi- sensor data fusion, variable speed operation, noisy measurement environments, and the requirement for real- time inference on edge computing hardware. This study presents a CNN- LSTM hybrid deep learning architecture, designated FaultNet- 5, designed for multi- class fault detection and severity classification in rotating machinery under variable speed conditions using simultaneous tri- axial accelerometer measurements. FaultNet- 5 achieved 97.8% overall classification accuracy and weighted F1 score of 0.973 on a held- out test set comprising 12,400 vibration records across five fault classes, outperforming Support Vector Machine (91.2%), Random Forest (94.8%), and gradient boosting (95.4%) baseline classifiers. Deployed on NVIDIA Jetson Xavier NX edge hardware, FaultNet- 5 achieves 4.2 ms per- inference latency, enabling real- time fault detection at data acquisition rates of 20 kHz. Industrial validation across three production facilities over nine months demonstrated 94.3% early fault detection rate with 8.6- day mean precursor detection lead time, enabling proactive maintenance interventions that reduced unplanned downtime by 73.4% and maintenance cost by 41.2%.