Predicting the Remaining Useful Life (RUL) of aircraft jet engines using sensor time-series data from the NASA CMAPSS dataset.

Aircraft engines generate continuous sensor data during operation. As components degrade over time, failure risk increases. Unplanned failures cause:

• 1. Safety risks — catastrophic failure mid-flight
• 2. High costs — emergency maintenance is 3–5x more expensive than scheduled
• 3. Downtime losses — airlines lose ~$150K/hour per grounded aircraft

Goal: Predict how many operational cycles remain before an engine fails — enabling maintenance before failure, not after.

RUL = Number of operational cycles remaining before engine failure

NASA CMAPSS (Commercial Modular Aero-Propulsion System Simulation)

26 columns per file: engine_id, cycle, operational_setting_1/2/3, sensor_1 to sensor_21

Download: https://www.nasa.gov/content/prognostics-center-of-excellence-data-set-repository

predictive-maintenance-rul/
│
├── notebooks/
│   └── predictive_maintenance_full.ipynb
│
├── data/
│   ├── train_FD001.txt
│   ├── train_FD002.txt
│   ├── train_FD003.txt
│   └── train_FD004.txt
│
├── models/
│
├── README.md
└── requirements.txt

Each of the 4 datasets goes through the same pipeline independently:

Load Dataset
    ↓
Create RUL (clipped at 125 cycles)
    ↓
EDA
    ↓
Feature Selection
    ↓
Rolling Mean Features
    ↓
Train/Test Split
    ↓
MinMax Scaling
    ↓
Random Forest + XGBoost
    ↓
Model Comparison

Then cross-dataset results are combined into a unified comparison table.

Raw RUL goes 300+ cycles for new engines. Model wastes capacity learning the perfectly healthy phase. Clipping at 125 focuses training on the degradation phase that actually matters.

Random row splitting causes data leakage. Instead, whole engines are assigned to train or test sets.

FD001/FD003 have one operating condition, so operational settings are constant and removed.

FD002/FD004 have multiple operating conditions, so settings are retained.

Rolling mean over 10 cycles per engine captures degradation trends and reduces sensor noise.

XGBoost outperforms Random Forest across all four datasets.

Best Overall Model: FD003 XGBoost RMSE = 13.14 R² = 0.897

An RMSE of approximately 18 cycles means predictions are accurate within about 18 flight cycles.

This provides roughly an 18-day early warning window for maintenance planning.

Estimated Savings: ~$400K per avoided emergency maintenance event.

SHAP (SHapley Additive exPlanations) is used to explain XGBoost predictions and identify which sensors contribute most to Remaining Useful Life estimation.

git clone https://github.com/your-username/predictive-maintenance-rul.git
cd predictive-maintenance-rul

pip install -r requirements.txt

jupyter notebook predictive_maintenance_full.ipynb

Run all cells from top to bottom.

pandas
numpy
matplotlib
seaborn
scikit-learn
xgboost
shap
joblib
jupyter

rf_FD001.pkl
xgb_FD001.pkl
scaler_FD001.pkl

rf_FD002.pkl
xgb_FD002.pkl
scaler_FD002.pkl

rf_FD003.pkl
xgb_FD003.pkl
scaler_FD003.pkl

rf_FD004.pkl
xgb_FD004.pkl
scaler_FD004.pkl

• Predictive Maintenance
• Machine Learning
• Feature Engineering
• Time Series Analysis
• Random Forest
• XGBoost
• SHAP Explainability
• Python
• Pandas
• Scikit-learn

Sanil J R

Data scientist