🛫 Aircraft Predictive Maintenance System

A Production-Grade Machine Learning System designed to predict aircraft engine failures before they happen. Built with XGBoost, FastAPI, and robust Time-Series Engineering principles.

---

🚀 Key Features

• Failure Prediction: Accurately predicts if an engine will fail within the next 20 cycles (Binary Classification).
• Leakage-Proof Validation: Implements Unit-Based Splitting to ensure the model generalizes to new, unseen engines.
• Time-Series Engineering: auto-calculates Rolling Statistics (Mean/Std) and Lag Features to capture degradation trends.
• Real-Time Inference: Exposes a high-performance FastAPI endpoint with <15ms latency.
• Metadata Tracking: Automatically versions models with training timestamps, hyperparameters, and evaluation metrics.

---

📊 Performance Metrics

We optimize for Recall (capturing as many failures as possible) while maintaining high Precision.

Metric	Score	Meaning
ROC-AUC	0.9967	Excellent discrimination between healthy and failing states.
Recall	0.7480	Accurately identifies ~75% of imminent failures (Safety Critical).
Precision	0.6619	~66% of alerts are real failures (Cost Effectiveness).
F1 Score	0.7023	Strong balance between safety and maintenance costs.
Threshold	0.28	Tuned threshold to prioritize safety over default 0.5.
Latency	~10ms	Inference time per request (99th percentile).

---

🏗️ System Architecture

The project follows a modular, clear separation of concerns:

graph LR
    A[Raw Sensor Data] -->|Cleaning| B(Data Processing)
    B -->|Feature Engineering| C{Feature Store}
    C -->|Rolling/Lag/Trend| D[XGBoost Classifier]
    D -->|Artifacts| E[Model Registry]
    E --> F[FastAPI Service]
    G[Client Request] --> F
    F -->|JSON Response| H[Failure Probability]

Loading

🧠 Model Strategy

• Algorithm: XGBoost Classifier (Gradient Boosting Trees).
  • Why? Proven champion for tabular/sensor data, handles non-linearities, robust to outliers.
• Target: label = 1 if RUL <= 20 else 0.
• Engineering:
  • Rolling Window (5): Smooths sensor noise (Mean) and captures volatility (Std).
  • Lags (t-1, t-2): Captures the "velocity" of degradation.

---

🛠️ Installation & Setup (Step-by-Step)

Prerequisites

• Python 3.10+
• pip and virtualenv

1. Clone & Setup

# Clone repository
git clone https://github.com/your-username/aircraft-predictive-maintenance.git
cd aircraft-predictive-maintenance

# Create Virtual Environment & Install Dependencies
# (Or use the helper command)
make setup

Alternatively: python3 -m venv venv && source venv/bin/activate && pip install -r requirements.txt

2. Data Processing

Fetches the NASA CMAPSS dataset and generates engineering features.

make process

Output: data/processed/train.csv (Enriched with 100+ features)

3. Train Model

Trains the XGBoost classifier, performs time-aware splitting, and saves artifacts.

make train

Output: models/xgb_model.joblib, models/metadata.json

4. Evaluate Performance

Generates a detailed markdown report with metrics, confusion matrix, and latency.

make evaluate

Output: eval/report.md

5. Start API Server

Launches the High-Performance FastAPI server.

make serve

Access Swagger UI: http://localhost:8000/docs

---

🔌 API Usage

Endpoint: POST /predict

Request Example

Send raw sensor data for the current cycle. The API handles feature scaling/mapping.

curl -X POST http://localhost:8000/predict \
-H "Content-Type: application/json" \
-d '{
  "sensor_1": 642.00,
  "sensor_2": 1590.00,
  "sensor_3": 1405.00,
  "cycle": 150
}'

Response Example

{
  "failure_probability": 0.824,
  "risk_level": "HIGH",
  "top_features": {
    "sensor_11_mean_5": 0.34,
    "sensor_4_mean_5": 0.12,
    "sensor_9_std_5": 0.08
  },
  "threshold_used": 0.28,
  "inference_ms": 13.103,
  "model_version": "2026-02-12T02:59:02..."
}

---

📂 Project Structure

aircraft-predictive-maintenance/
├── data/               # Raw and processed datasets
├── models/             # Serialized models and metadata
├── notebooks/          # Exploratory analysis
├── src/                # Source code
│   ├── api/            # FastAPI application
│   ├── features/       # Data cleaning & feature engineering
│   ├── training/       # Model training loop
│   └── evaluation/     # Metrics & reporting
├── eval/               # Generated reports
├── Makefile            # Automation shortcuts
└── requirements.txt    # Dependencies

---

⚠️ Comparison to "Real World"

This is a sophisticated demo. A full production version would add:

1. Feature Store (Feast/Tecton): To serve historical rolling features in real-time instead of the API assuming steady-state.
2. Model Registry (MLflow): For deeper experiment tracking.
3. CI/CD: Automatic retraining pipelines on new data ingestion.

---

Author: Ritesh License: MIT