This project develops an end-to-end predictive maintenance pipeline to estimate the Remaining Useful Life (RUL) of turbofan engines using the NASA CMAPSS dataset.

The goal is to demonstrate:

• Sound exploratory data analysis (EDA)
• Degradation-aware feature engineering
• Leakage-safe model evaluation
• Defensible model selection based on generalization and interpretability

The project uses the FD001 subset, which represents a single operating condition and a single fault mode.

• Understand engine degradation behavior from multivariate time-series sensor data
• Engineer features that capture progressive damage accumulation
• Compare linear and nonlinear regression models for RUL prediction
• Select a robust baseline model suitable for real-world maintenance decision-making
• Provide interpretable insights into model behavior

• Source: NASA CMAPSS (Commercial Modular Aero-Propulsion System Simulation)
• Subset: FD001
• Engines: 100 training engines, run-to-failure
• Sensors: 21 sensor measurements + operational settings
• Target: Remaining Useful Life (RUL), computed from run-to-failure trajectories

NASA-CMAPSS-Predictive-Maintenance/
│
├── data/
│   ├── raw/                      # Original CMAPSS files
│   └── processed/                # Cleaned & engineered datasets
│
├── notebooks/
│   ├── 01_eda.ipynb
│   ├── 02_feature_engineering.ipynb
│   └── 03_modeling.ipynb
│
├── src/
│   ├── features.py               # Reusable feature functions
│   ├── models.py
│   └── utils.py
│
├── results/
│   ├── figures/                  # Plots and visualizations
│   ├── metrics.csv               # Final model metrics
│   └── ridge_final.joblib        # Saved final model
│
├── requirements.txt
└── README.md

Notebook: 01_eda.ipynb

Key findings from EDA:

• Engine lifetimes vary significantly, requiring engine-specific modeling
• Several sensors exhibit monotonic degradation trends as failure approaches
• Multiple sensors show strong correlation with RUL
• A subset of sensors has near-zero variance and provides no predictive value
• Operational settings have minimal influence in FD001
• Degradation behavior is progressive rather than abrupt

These insights guided feature selection and engineering strategies.

Notebook: 02_feature_engineering.ipynb

Feature engineering focuses on capturing degradation dynamics rather than raw sensor values.

Key feature groups:

• Per-engine standardization – Emphasizes relative degradation patterns instead of absolute sensor levels
• Cumulative damage indicators – Approximate irreversible wear accumulation over time
• Rolling statistics & slopes – Capture short-term variability and long-term degradation trends
• Rate-of-change features – Identify accelerating degradation behavior
• Limited interaction terms – Introduced only for highly correlated sensor pairs
• Normalized lifecycle position (cycle_norm) – Represents relative engine life progression

Feature selection is performed using mutual information on engine endpoints to avoid temporal leakage.

Notebook: 03_modeling.ipynb

Models are evaluated progressively:

1. Ridge Regression (baseline) – Linear model with L2 regularization, strong stability, high interpretability
2. XGBoost – Nonlinear, tree-based ensemble with high training capacity
3. MLP (Neural Network) – High-capacity nonlinear model

• Train–validation split performed by engine ID (no leakage across cycles)
• Test evaluation uses only the last cycle per engine
• Metrics: RMSE, MAE, R²
• Test performance prioritized over training metrics

While XGBoost and MLP achieve excellent training and validation performance, they fail to generalize to unseen test engines due to overfitting to engine-specific degradation trajectories.

Ridge regression generalizes substantially better, making it the most reliable model for deployment.

SHAP (SHapley Additive exPlanations) is used to analyze feature contributions.

• XGBoost SHAP shows heavy reliance on lifecycle proxy features (e.g., normalized cycle position), explaining poor generalization
• Ridge SHAP reveals more evenly distributed, physically meaningful degradation features, supporting its robustness

To assess real-world usefulness, RUL predictions are converted into a binary classification task using a 30-cycle threshold.

Ridge Regression Performance (Test Set):

• Precision: 0.96
• Recall: 0.92
• F1-score: 0.94

This demonstrates strong capability to identify engines approaching failure.

Ridge Regression is selected as the final model due to:

• Superior test generalization
• Stability in high-dimensional feature space
• Interpretability of degradation drivers
• Practical maintenance decision performance

The trained model is saved for reproducibility and downstream use.

• Simpler models can outperform complex models in low-sample, high-dimensional settings
• Engine-level data leakage prevention is critical for honest evaluation
• Degradation-aware features are more important than model complexity
• Interpretability is a strength, not a trade-off, in predictive maintenance

• Python, NumPy, Pandas
• Scikit-learn
• XGBoost
• TensorFlow / Keras
• SHAP
• Matplotlib, Seaborn