Early-Warning Dashboard for Plant Sensors
A FastAPI and React dashboard that learns each sensor's normal behaviour and warns before limits are crossed, using Monte-Carlo, Chronos and LSTM forecasts.
- Role
- Working-student project at Fraunhofer IPK.
- Forecasts
- Monte Carlo, Chronos, LSTM
- Modes
- Live + batch replay
- Stack
- Python, FastAPI, WebSockets, React, Vite, pandas / NumPy, Numba, PyTorch, Chronos, HDF5, MySQL
What it does
An early-warning system for industrial sensor data. Every signal runs through its own state machine (start-up, inactive, calibration, monitoring): it learns its normal level during calibration, then raises alarms on deviations and forecast warnings when a Monte-Carlo projection, the Chronos time-series model or an LSTM predicts that a threshold will be exceeded. A live mode polls a MySQL database; a batch mode replays historical HDF5 recordings and exports plots and data, to check how early a known fault would have been detected. Results stream to a React dashboard over WebSockets.
- Data
- MySQL
- HDF5
- Engine
- State machine
- Forecasts
- API
- FastAPI
- WebSocket
- UI
- React
The hard part
Replaying long recordings fast while keeping results reproducible: incremental state-machine steps, a JIT-compiled fast path and seeded forecasts, so identical runs give identical warnings.