Predictive Maintenance — Aircraft Turbofan Engine RUL Prediction

Predicts the Remaining Useful Life (RUL) of aircraft turbofan engines using sensor and operational data from the NASA C-MAPSS benchmark dataset. Includes a Plotly Dash dashboard for interactive exploration of sensor degradation and model performance.

---

Model Performance

Best model: LightGBM + first-n-cycle normalisation + RUL clipping

Split	MAE	RMSE	R²
Test (80/20 split)	7.55 cycles	11.27 cycles	0.92
Validation (held-out NASA test set)	11.14 cycles	15.28 cycles	0.86

Four preprocessing strategies were compared systematically — see the notebook for the full comparison.

---

Project Structure

Predictive-Maintenance/
├── dataset/                     # C-MAPSS train/test/RUL files
├── models/                      # Saved LightGBM model files (.joblib)
├── predictive_maintenance.ipynb # Full pipeline: EDA → feature engineering → modelling
├── app.py                       # Plotly Dash interactive dashboard
├── requirements.txt
└── README.md

---

Dataset: NASA C-MAPSS FD001

100 turbofan engines simulated from healthy state to failure. Each row is one engine cycle.

Column group	Details
Engine ID + cycle	Unique engine identifier and time step
Operational settings	3 columns (altitude, Mach number, throttle)
Sensor readings	21 sensors (temperatures, pressures, speeds, flows)
RUL (train only)	Calculated as max_cycle − current_cycle

7 constant sensors (s1, s5, s6, s10, s16, s18, s19) were identified and dropped.

---

Feature Engineering

Technique	Purpose
Min-Max normalisation	Baseline sensor scaling
First-n-cycle normalisation	Engine-relative scaling — eliminates inter-engine bias
RUL clipping (max = 121)	Reduces noise from early-life cycles, improves convergence
Rolling averages	Smooths sensor signal noise
Variance threshold	Removes constant/near-constant features

The comparison across 4 preprocessing strategies shows that first-n-cycle normalisation + RUL clipping generalises best to unseen engines (R² 0.86 vs 0.20 for the baseline).

---

Setup

git clone https://github.com/KonulJ/Predictive-Maintenance.git
cd Predictive-Maintenance
pip install -r requirements.txt

Download the C-MAPSS FD001 dataset from NASA Prognostics Center and place the .txt files in dataset/.

python app.py

The Dash dashboard opens at http://localhost:8050.

---

Key Concepts Demonstrated

Concept	Implementation
Regression on multivariate time series	LightGBM + XGBoost on 21-sensor sequential data
Feature engineering strategy comparison	4 normalisation + clipping combinations, evaluated on held-out set
Overfitting diagnosis	Baseline R² 0.91 (test) → −0.61 (validation); fixed by first-n-cycle normalisation
Hyperparameter tuning	GridSearchCV with 5-fold CV across 81 parameter combinations
Industrial ML context	NASA benchmark, condition-based maintenance framing

---

Roadmap

• Classification model for early-warning risk zones (within 30 cycles of failure)
• LSTM for sequence-aware RUL prediction
• Docker deployment of the Dash dashboard

---

Built by Konul Jafarova