PulseNet - Remaining Useful Life Prediction for Turbofan Engines

PulseNet predicts how many operating cycles a turbofan engine has left before failure, using NASA's C-MAPSS dataset (FD001 subset). It includes a training pipeline, a FastAPI inference server, and a Streamlit dashboard.

What It Does

• Trains models on NASA C-MAPSS FD001 run-to-failure data (100 training engines, 100 test engines)
• Predicts Remaining Useful Life (RUL) at any point in an engine's operating history
• Serves predictions via a REST API with JWT authentication
• Provides a monitoring dashboard (Streamlit)

What It Cannot Do

• It only works with C-MAPSS-format sensor data (21 sensors, 3 operational settings). It won't generalize to arbitrary industrial equipment without retraining on new data.
• The anomaly detection model (Isolation Forest) has limited precision. On the FD001 benchmark it achieves F1 = 0.54, Precision = 0.71, Recall = 0.43 — it catches most failures but at the cost of false alarms. See the Benchmark Results section for the full validated numbers.
• The RUL regression model (Random Forest) achieves RMSE in the 15–25 cycle range on FD001. This range is a test-assertion bound from the CI test suite, not a single point estimate — the exact value varies by run. A typical classical baseline on FD001 falls in this range.
• This is a reference implementation, not a production-certified system. It has not been validated on real-world fleet data.

Models

Model	Type	What It Predicts
Isolation Forest	Anomaly detection	Binary: "degrading" vs "healthy"
Random Forest	Regression	RUL in cycles (0-125, capped)
LSTM	Deep learning (sequence)	RUL in cycles
Transformer	Deep learning (attention)	RUL in cycles

The default model is Isolation Forest for anomaly detection. The RUL regression module uses Random Forest with rolling-mean features.

Benchmark Results (FD001)

From results/validation_results.json (ran on official NASA C-MAPSS data):

Isolation Forest (anomaly detection):

• F1: 0.54, Precision: 0.71, Recall: 0.43, ROC-AUC: 0.70
• Training time: 0.26 seconds on 45 features, 20,631 training rows

RUL Regression (Random Forest, official per-unit split):

• RMSE: 15–25 cycles (test-assertion range; CI asserts the result falls within this bound — the exact value depends on the run and is not a single pinned benchmark number)
• Uses the C-MAPSS asymmetric scoring function (late predictions penalised more than early ones)
• No random splitting — train and test engines are disjoint, as intended by NASA

Inference latency (Isolation Forest):

• Mean: 2.7ms, P99: 4.3ms (500 samples)
• Throughput: ~13,400 samples/sec at batch size 32

Requirements

• Python 3.10+
• ~250 MB RAM per instance
• GPU optional (only used for LSTM/Transformer training)

Install

git clone https://github.com/poojakira/PulseNet-RUL-Forecasting
cd PulseNet-RUL-Forecasting

# Create a virtual environment
python -m venv .venv
.venv\Scripts\activate        # Windows
# source .venv/bin/activate   # Linux/Mac

# Install dependencies
pip install -e ".[dev]"

Download the Data

python scripts/download_data.py

This downloads the NASA C-MAPSS dataset (~12 MB zip) to data/official/ and verifies its SHA-256 hash.

Run the Pipeline

# Full pipeline: ingest → preprocess → train → evaluate
python main_pipeline.py --mode full

# Just train
python main_pipeline.py --mode train

# Run predictions
python main_pipeline.py --mode predict

# Run benchmarks
python main_pipeline.py --mode benchmark

Start the API Server

# Copy and edit environment config
copy .env.example .env
# Fill in PULSENET_JWT_SECRET and PULSENET_ENCRYPTION_KEY

# Start the server
uvicorn pulsenet.api.app:app --host 0.0.0.0 --port 8000

API docs at http://localhost:8000/docs once running.

Start the Dashboard

streamlit run src/pulsenet/dashboard/app.py

Opens at http://localhost:8501.

Run Tests

pytest tests/ -v --cov=pulsenet --cov-report=term-missing

Tests require 80% coverage to pass (configured in pyproject.toml).

Docker

docker build -t pulsenet:latest .
docker-compose up -d

Project Structure

src/pulsenet/
  api/          - FastAPI server with JWT auth, prediction and health routes
  dashboard/    - Streamlit monitoring dashboard
  evaluation/   - RUL regression metrics (RMSE, C-MAPSS score)
  models/       - Isolation Forest, LSTM, Transformer, Ensemble
  pipeline/     - Data ingestion, preprocessing, orchestration
  streaming/    - Async producer/consumer for real-time inference
  security/     - Encryption, audit logging
  mlops/        - MLflow experiment tracking

Dataset

NASA C-MAPSS FD001: Simulated turbofan engine degradation data.

• 100 training engines (run to failure)
• 100 test engines (truncated before failure)
• 21 sensor channels + 3 operational settings
• Source: https://data.nasa.gov/dataset/cmapss-jet-engine-simulated-data

License

Apache-2.0