← Week 09

Intro to ML · Time series

The pump investigation

Eight connected experiments follow Serega from sensor readings to an alarm decision. Open a scene below, or present the full lectures.

Lecture 1 · HTML · Lecture 2 · HTML
Lecture 1 · PDF with keyframes · Lecture 2 · PDF with keyframes

Learning targets and reference details · Forecast-quality report

SceneQuestionExplore
1Does the pump repeat itself?
Shifted copies, lag plots and lag correlation.
Lag and selected pair.
2What changed in the readings?
A moving window, mean and standard deviation.
Observation time and window length.
3Which rhythms are present?
Wave components, spectrum and reconstruction.
Components and phase.
4Is a change the same as a fault?
A sustained shift and a healthy reference.
Time and adaptive window.
5Scores rank; thresholds decide
Precision, recall, F1 and AP on a small timeline.
Threshold and ranking.
6Can a filter extract a useful feature?
Temporal filtering, then spatial mixing.
Temporal position and spatial weight.
7Has the fault ended?
A stored forecast becomes accurate inside a fault.
Time and observation availability.
8When should Serega raise an alarm?
Real validation scores and an alarm budget.
ECDF cutoff.

Present step by step

→ / Space advances a scene, then moves to the next slide. ← goes back. Shift + arrows moves directly between slides. The on-slide Back, Next step and Restart buttons do the same scene operations.

Investigate independently

Choose Explore, or press E, to reveal the sliders. Every input recomputes the plotted result. Returning to Lecture resumes the previous lecture stop. R restarts the current mode. ? shows all shortcuts.

Scenes 1–7 use explicitly synthetic or illustrative signals and fixed teaching weights. Scene 8 uses healthy calibration and validation data from the recorded SKAB experiment; held-out test results are reserved for the lab. All scenes work offline. No model download or live inference is required. The PDF versions include three selected frames per scene.