Lab 9 · When should the pump stay in alarm?
90 minutes · Student notebooks in English and Russian · Free Colab CPU
Use real SKAB pump measurements to detect the entire annotated faulty period. Compare robust statistics, PCA, Isolation Forest, classical score fusion and forecasting residuals from Chronos-2 and a matched persistence baseline. Fault onset is a secondary target.
Open your student notebook
Save your own copy in Drive before editing. If Colab cannot open a link, download the notebook here and use File → Upload notebook in Colab. You do not need access to the course’s GitHub repository.
The experimental route
- Define the fault-state target and split whole runs before constructing windows.
- Compare current measurements with causal rolling features while keeping the model family fixed.
- Calibrate detector scores on a separate normal reference and combine complementary sources.
- Compare Chronos-2 and persistence forecasts, then distinguish forecast error from fault-state coverage.
- Choose a pipeline and alarm threshold using validation only; freeze that decision before revealing held-out runs.
- Explain the result using timelines, ranking metrics, state coverage and false alarms.
Four small code tasks cover PCA residual distance, causal rolling features, empirical score calibration and weighted score fusion. Work section by section: an unfinished task deliberately raises NotImplementedError. Complete the task before continuing.
Runtime and supplied assets
The required route uses CPU and includes the pinned data, inspectable helpers and genuine recorded Chronos-2 forecasts. Classical models are fitted during your run. No API key, GPU, model download or paid Colab plan is required.
An optional extension runs a new foundation-model forecast and downloads model weights. GPU can speed up that extension; it is not required for the laboratory.
Before and after the lab
Two Week 9 lectures and PDF slides · Interactive pump scenes · Learning guide
Use measured evidence to defend a decision rather than trying to reproduce a predetermined model ordering. This is a small run-held-out study on one installation, not evidence of general superiority for a model family.
Data: SKAB, Katser & Kozitsin (2020), dataset DOI, GPL-3.0. Exact data and model revisions are documented inside the notebooks.