Catalyst Engine Monitor

We predict turbofan engine failures before they happen by streaming 21 live sensors through Kafka and using an LSTM to spot early signs of degradation in real time.

Comment

Inspiration

Unplanned engine failures cost the aviation industry over $150 billion annually. Current maintenance approaches are reactive—fix it when it breaks. With modern sensors generating thousands of data points, we saw an opportunity to flip this model: predict failure risk in real-time and intervene proactively.

What it does

Catalyst is a real-time predictive maintenance system for turbofan engines:

  1. Streams sensor data — 21 sensors measuring temperature, pressure, and speed flow through Confluent Kafka
  2. Computes features — Rolling statistics detect degradation patterns in real-time
  3. Predicts risk — LSTM neural network scores failure risk from 0-100%
  4. Enables intervention — Operators can simulate maintenance actions and see cycles saved
  5. Explains alerts — Gemini AI generates natural language explanations

How we built it

Streaming Layer:

  • Confluent Cloud with 3 Kafka topics (sensor-events, feature-events, risk-events)
  • WebSocket server for real-time dashboard updates

ML Pipeline:

  • LSTM neural network trained on NASA C-MAPSS FD001 dataset (100 engines, 21 sensors)
  • Multi-task learning: predicts both risk score and remaining useful life
  • Physics-informed features: rolling variance, trend slopes, cross-sensor correlations

Cloud Infrastructure:

  • Google Cloud Run for serverless backend
  • Vertex AI for ML model serving with auto-scaling
  • Gemini 2.0 Flash for AI-powered alert explanations

Frontend:

  • React 18 with real-time WebSocket updates
  • Recharts for live sensor visualization
  • Fleet and single-engine views with maintenance simulation

Challenges we ran into

  • Progressive windowing: Features need to work from cycle 2, not just after a full 30-cycle window
  • Multi-sensor fusion: Combining 14 useful sensors into meaningful health indicators
  • Real-time constraints: Keeping prediction latency under 100ms while computing rolling stats
  • Intervention modeling: Simulating how maintenance actions affect degradation trajectories

Accomplishments we're proud of

  • 94% failure detection rate with 25-cycle average lead time
  • End-to-end streaming pipeline from sensor to dashboard in under 200ms
  • Maintenance simulation that shows tangible cycles saved
  • Clean architecture with clear separation between streaming, ML, and presentation layers

What we learned

  • Confluent Kafka's exactly-once semantics are crucial for ML predictions
  • Google Cloud Run handles WebSocket connections surprisingly well
  • Physics-informed features (variance, slopes) often outperform raw sensor values
  • Real-time ML is more about feature engineering than model complexity

What's next for Catalyst

  • Multi-engine correlation: Detect fleet-wide issues affecting multiple engines
  • Feedback loop: Track actual failures vs predictions to retrain models
  • Mobile alerts: Push notifications for critical risk thresholds
  • Schema Registry: Enforce data contracts across the streaming pipeline

Built With

Share this project:

Updates