Deep Learning-Based Predictive Maintenance and Remaining Useful Life Estimation of Turbofan Engines Using CNN-BiLSTM Multi-Task Learning

This study develops a deep learning-based predictive maintenance framework for Remaining Useful Life (RUL) estimation and failure-risk assessment of turbofan engines using the NASA C-MAPSS FD001 dataset.

The proposed framework uses a CNN-BiLSTM architecture within a multi-task learning setup. Sixteen selected sensor features are standardized using training-data statistics and organized into 30-cycle temporal windows. The model performs two tasks: continuous RUL regression and failure-risk classification.

The RUL regression component achieved an MAE of 28.9519 cycles, RMSE of 39.2669 cycles, and R² of 0.4825. The failure-risk component achieved a PR-AUC of 0.9183, ROC-AUC of 0.9966, precision of 0.7944, recall of 0.8494, and F1-score of 0.8210. At the engine level, the framework achieved 88% warning coverage, a mean warning lead time of 30.91 cycles, and 3 missed failures out of 25 evaluated failures.

The trained model was integrated into a Streamlit dashboard to demonstrate RUL estimation, failure-risk assessment, and maintenance-oriented status recommendations. The study highlights the potential of temporal deep learning and multi-task learning for predictive maintenance while acknowledging the limitations of benchmark simulation data and the uncertainty of RUL prediction.

Keywords: Predictive Maintenance, Remaining Useful Life, NASA C-MAPSS, CNN, BiLSTM, Multi-Task Learning, Deep Learning, Regression Analysis.