Fault detection in a chemical process plant

The brief

A chemical plant has dozens of sensors, and a fault can show in several at once, slowly. Use the Tennessee Eastman process data to build a model that detects a fault early and says which kind it is, and compare simple statistical monitoring with machine learning. Good looks like: detection delay measured for each fault, false alarms counted on normal runs, and the faults that no method catches named.

What you will build

1. Load the normal and faulty runs
2. Build PCA-based monitoring as a baseline
3. Train a classifier to name the fault
4. Measure detection delay and false alarm rate
5. Compare the methods by fault type
6. Write a short guide for operators

Free, open data to build it with