LSTM-CNN Hybrid Deep Learning Architecture for Predictive Maintenance and Remaining Useful Life Estimation in Rotating Industrial Machinery

Unplanned downtime of rotating machinery — including motors, pumps, compressors, and gearboxes — in Indian manufacturing plants accounts for an estimated 8-12% of production time lost annually, costing the sector over ₹22,000 crore per year according to the Confederation of Indian Industry's 2023 Maintenance Benchmarking Survey. Traditional time-based preventive maintenance schedules, calibrated to average failure rates rather than individual machine condition, both over-maintain machines in good health and fail to prevent the minority of accelerated-wear events that cause most unplanned stoppages. Condition-Based Maintenance (CBM) driven by real-time vibration, current, and temperature signals offers the theoretical promise of maintenance precisely at the point of need — but realising this promise requires fault classification and Remaining Useful Life (RUL) estimation algorithms accurate enough to generate actionable maintenance alerts ahead of critical failure. This paper presents a hybrid LSTM-CNN architecture that processes raw vibration time-series through one-dimensional convolutional layers for local feature extraction and bidirectional LSTM layers for temporal dependency modelling, enabling simultaneous multi-class fault diagnosis (normal, bearing outer race fault, bearing inner race fault, gear tooth fracture) and continuous RUL estimation. The model is trained on the CWRU Bearing Dataset supplemented by data from a purpose-built test rig at IIT Bombay that introduces progressive bearing degradation under controlled load and speed conditions representative of Indian industrial duty cycles. Classification accuracy of 97.8% (macro-F1), AUC of 0.991, and RUL prediction RMSE of 38.4 cycles demonstrate performance superior to SVM, Decision Tree, Random Forest, CNN-1D, and BiLSTM baselines. The TU Munich collaboration contributes transfer learning validation on BMW Group gearbox data, confirming cross-domain applicability