✈️ NASA CMAPSS Turbofan Engine — Predictive Maintenance System

🚀 Detects engine failure 122 cycles early with explainable AI — deployed as a production-grade Streamlit dashboard.

---

📸 Dashboard Preview

🔧 Engine Status Panel

Real-time monitoring of engine health indicators and anomaly detection outputs.

🧠 XAI Sensor Analysis

Heatmap visualization explaining which sensors contributed most to detected anomalies.

Engine Status Panel	XAI Sensor Analysis

---

📋 Table of Contents

• Why This Project Matters
• Key Results
• Project Overview
• How It Works
• Dashboard Features
• Project Structure
• Installation & Running
• Tools & Versions
• Key Learnings
• Limitations
• Future Work
• Demo
• Dataset
• Author

---

💡 Why This Project Matters

Unplanned aircraft engine failures are among the most costly and dangerous events in commercial aviation. Traditional time-based maintenance services every engine on a fixed schedule regardless of actual condition — leading to wasted resources or, worse, missed failures.

This system addresses that gap by:

• 🔍 Detecting degradation 122 cycles before failure — giving maintenance teams a large, actionable window
• 💰 Reducing maintenance costs by an estimated 96% — from ~$2M to ~$80K on a 100-engine fleet
• 🔬 Explaining every alert — XAI identifies the exact failing sensor (s14, s9, or s11), not just a black-box flag
• 🚀 Deploying as a production web app — usable by non-technical maintenance operators with no data science background

Predictive maintenance powered by interpretable AI can save millions and prevent accidents in real-world aviation systems.

---

📊 Key Results

Metric	Value	Source
🏆 Best ROC-AUC	0.7811	Fuzzy Clustering (Notebook baseline)
🤖 LSTM ROC-AUC	0.6915	LSTM Autoencoder (Notebook baseline)
🔍 Silhouette Score	0.6433	Fuzzy Clustering — strong cluster separation
✅ Engine Detection Rate	100% (20/20)	All baseline models
⏱️ Early Warning Lead Time	122 cycles	Before failure point
🔬 Primary Failure Sensors	s14, s9, s11	XAI per-sensor MSE decomposition
💸 Estimated Cost Reduction	96%	$2M → $80K on 100-engine fleet

---

🎯 Project Overview

This project follows a dual-phase architecture that mirrors professional data science practice — separating exploratory research from production engineering.

Phase	Deliverable	Focus
Phase 1 — Research	notebooks/1-EDA_and_Cleaning.ipynb	EDA, global IQR exploration, MinMaxScaler, LSTM Autoencoder & Fuzzy Clustering baselines
Phase 2 — Production	app/app.py	PCA Reconstruction Autoencoder, per-engine IQR, StandardScaler, XAI, RUL forecasting, fleet batch report

Why the two phases differ by design:

Decision	Notebook (Research)	App (Production)
IQR method	Global — across all engines	Per-engine — respects each engine's baseline
Outlier handling	Flags created, no rows removed	Rows with ≥3 flags removed
Scaler	MinMaxScaler [0,1] — required for LSTM	StandardScaler — required for PCA distance

---

⚙️ How It Works

Production Pipeline (app/app.py)

Raw Data (20,631 rows × 26 cols)
        ↓
  7-Step Cleaning
  (per-engine IQR · drop 7 constant columns · remove rows with ≥3 outlier flags)
        ↓
  Feature Engineering
  (RUL = max_cycle − current_cycle · 13 STRONG_FEATURES with |r| > 0.5)
        ↓
  StandardScaler → fit on healthy engines only (RUL > 125)
        ↓
  PCA Reconstruction Autoencoder (95% variance retained)
        ↓
  Anomaly Score = Mean MSE over last 50 cycles
        ↓
  Percentile Thresholds — derived from healthy training distribution
  (95th → MONITOR · 98th → WARNING · 99.5th → URGENT)
        ↓
  ┌──────────────────────────────────────────────────────┐
  │  🟢 HEALTHY · 🟡 MONITOR · 🟠 WARNING · 🔴 URGENT  │
  └──────────────────────────────────────────────────────┘
        ↓
  XAI: Per-sensor MSE decomposition
  → s14 (+262%) · s9 (+193%) · s11 (+81%)

Anomaly Threshold Derivation

All four zone boundaries are statistically derived from training data — no hardcoded values:

Zone	Boundary	Operational Action
🟢 HEALTHY	< 95th percentile MSE	Normal operations
🟡 MONITOR	95th – 98th percentile	Inspect within 30 days
🟠 WARNING	98th – 99.5th percentile	Inspect within 7 days
🔴 URGENT	> 99.5th percentile	Immediate maintenance required

---

🖥️ Dashboard Features

Tab	What You'll Find
📊 EDA	Engine lifetime distribution, sensor correlation heatmap, raw data sample, unique value audit
🧹 Data Cleaning	Cleaning log · per-engine IQR outlier chart · StandardScaler parameters · PCA summary cards · training error histogram proving threshold derivation
📉 RUL Analysis	Feature-RUL Pearson correlations · 6 sensor degradation subplots · Predicted vs Actual RUL with out-of-sample R² (leave-one-out)
🔬 Sensor XAI	Original vs PCA-reconstructed plots for s14, s9, s11 · per-sensor MSE heatmap · rolling error timeline showing exact failure crossing point
🚨 Batch Report	Fleet-wide status cards · urgent alert banners · sortable colour-coded engine table · CSV export

---

📁 Project Structure

NASA-Turbofan-Predictive-Maintenance/
│
├── app/
│   └── app.py                        # ✅ Production Streamlit dashboard
│
├── notebooks/
│   └── 1-EDA_and_Cleaning.ipynb      # 📓 EDA, cleaning, LSTM, Fuzzy baselines
│
├── assets/
│   ├── dashboard.png                 # 📸 Add screenshot after running the app
│   └── xai_heatmap.png               # 📸 Add XAI tab screenshot after running
│
├── data/
│   └── raw/
│       └── CMap_FD001/
│           └── train_FD001.txt       # 📊 NASA CMAPSS FD001 (download separately)
│
├── models/
│   └── scaler.pkl                    # 💾 MinMaxScaler from notebook (LSTM only)
│
├── reports/
│   └── Project_Report_final.docx     # 📄 Full 17-section technical report
│
├── requirements.txt                  # 📦 Python dependencies
├── .gitignore                        # 🚫 Files excluded from version control
├── LICENSE                           # ⚖️ MIT License
└── README.md                         # 📖 This file

---

🚀 Installation & Running

1. Clone the repository

git clone https://github.com/YOUR_USERNAME/NASA-Turbofan-Predictive-Maintenance.git
cd NASA-Turbofan-Predictive-Maintenance

2. Install dependencies

pip install -r requirements.txt

3. Launch the dashboard

streamlit run app/app.py

4. Use the dashboard

• Select "Use sample (FD001)" in the sidebar for an instant demo — no data download required
• Or upload your own train_FD001.txt for real results
• Pick any engine from the dropdown menu
• Navigate the 5 tabs: EDA → Data Cleaning → RUL Analysis → Sensor XAI → Batch Report
• Download the full fleet health report as CSV from Tab 5

---

🛠️ Tools & Versions

Production Dashboard (app/app.py)

Library	Version	Purpose
Streamlit	1.32+	Interactive web dashboard (5 tabs, CSV export)
Scikit-learn	1.4+	StandardScaler · PCA · LinearRegression
Pandas	2.0+	Data manipulation, per-engine groupby operations
NumPy	1.26+	Vectorised MSE computation, percentile thresholds
Matplotlib	3.8+	Custom sensor plots, XAI visualisations
Seaborn	0.13+	Correlation and MSE heatmaps

Research Notebook (notebooks/)

Library	Version	Purpose
PyTorch	2.0+	LSTM Autoencoder — custom class, Adam optimiser, MSELoss
scikit-fuzzy	0.4.2	Fuzzy C-Means clustering baseline
Statsmodels	0.14+	VIF multicollinearity analysis
Scikit-learn	1.4+	MinMaxScaler for LSTM input normalisation
Joblib	1.3+	Save/load fitted scaler as .pkl

---

🧠 Key Learnings

These reflect real engineering decisions made during the project — not just theory:

1. Transitioning research to production is the hardest part. The notebook and the app use different pipelines by design. MinMaxScaler works for LSTM because neural networks need bounded inputs. StandardScaler is required for PCA because reconstruction error is a Euclidean distance metric sensitive to variance, not range. Both are correct — for their context.

2. Data leakage hides in unexpected places. The RUL regression model uses leave-one-out: the selected engine is excluded from training entirely. A naive row-level train/test split would let the model see the same engine's earlier cycles during training and report an inflated R² that doesn't reflect real generalisation.

3. Outliers in predictive maintenance are not noise — they are signal. Global IQR incorrectly flags sensors that are normal for that engine but unusual compared to a different engine's baseline. Per-engine IQR was the only statistically defensible choice for production because each engine operates at slightly different baseline conditions.

4. Interpretability is as important as accuracy. A maintenance engineer needs to know which component is failing and when it crossed the threshold — not a probability score. The per-sensor MSE heatmap and rolling error timeline directly answer both questions, which a black-box model cannot.

5. Thresholds must come from data, not assumptions. Deriving zone boundaries from the 95th, 98th, and 99.5th percentiles of the healthy training distribution means the system self-calibrates to each dataset. Hardcoded thresholds break silently when the data distribution shifts.

---

⚠️ Limitations

Single operating condition only. The model was trained and evaluated on FD001 (one flight regime). Performance on FD002–FD004 (multiple operating conditions) has not been validated.

PCA trades accuracy for interpretability. PCA is a linear method and will not capture non-linear degradation patterns as effectively as a trained LSTM Autoencoder. The ROC-AUC gap (0.69 LSTM vs PCA proxy) reflects this trade-off — accepted because interpretability is essential for production deployment.

No real-time data streaming. The dashboard processes historical batch data. A production deployment would require integration with a live sensor feed and incremental model updates.

Synthetic sample data differs from real NASA data. The built-in demo uses synthetic data generated with realistic degradation trends. Exact threshold values and R² scores will differ when using the real dataset.

RUL forecasting is a linear baseline. LinearRegression on the PCA latent space is a first-order approximation. A dedicated LSTM regressor would capture the non-linear degradation curve more accurately.

---

🔭 Future Work

Priority	Enhancement	Why It Matters
🔴 High	Integrate trained LSTM .pth model file into the app	Replace PCA proxy with the trained deep learning model
🔴 High	Extend to multi-condition datasets (FD002–FD004)	Validate generalisation across operating regimes
🟡 Medium	Rolling re-training / online learning	Adapt to new data without full retraining
🟡 Medium	Adaptive percentile thresholds	Auto-adjust zone boundaries as data accumulates
🟡 Medium	Email / SMS alert integration for URGENT events	Notify maintenance teams automatically
🟢 Low	Deploy to Streamlit Cloud / AWS / Azure	Make the dashboard publicly accessible
🟢 Low	Hybrid model — deep learning + statistical methods	Combine LSTM temporal accuracy with PCA interpretability

---

🎥 Demo

[]

To record a demo:

1. Run streamlit run app/app.py
2. Screen-record a walkthrough of all 5 tabs
3. Upload to YouTube or Loom
4. Replace the placeholder below with your link

🎥 Watch Demo: [link_here]

---

📄 Technical Report

A full 17-section detailed technical report is included for in-depth analysis:

📥 Download: Technical Report

📚 Dataset

NASA CMAPSS FD001 — Commercial Modular Aero-Propulsion System Simulation

Saxena, A., Goebel, K., Simon, D., & Eklund, N. (2008). Damage Propagation Modeling for Aircraft Engine Run-to-Failure Simulation. NASA Ames Research Center.

Property	Value
Rows	20,631
Columns	26 (2 ID + 3 settings + 21 sensors)
Engines	100 (run-to-failure)
Operating condition	Single (FD001)
Headers / target variable	None — both constructed manually

📥 Download: https://ti.arc.nasa.gov/tech/dash/groups/pcoe/prognostic-data-repository/

Place the downloaded file at: data/raw/CMap_FD001/train_FD001.txt

---

👩‍💻 Author

Mennatullah Mohammed Kh
AI & Intelligent Systems Student — 3rd Year Computer Science - AI Engineer

🔗 LinkedIn: https://www.linkedin.com/in/menna-mohammed11/
🔗 GitHub: http://github.com/mennamohammedkh

Open to AI/ML internship and junior data scientist opportunities.

---

🏷️ Topics / Hashtags

LSTM Autoencoder Predictive Maintenance Anomaly Detection NASA CMAPSS RUL Estimation Fuzzy Clustering XAI Time Series PyTorch Streamlit Machine Learning Deep Learning

---

📄 License

MIT License — see LICENSE for details.

---

If this project was useful, consider giving it a ⭐ — it helps others find it.