Problem
Industrial teams monitor equipment with hundreds of sensors and almost no labelled failures. When a machine breaks, they must say when the state changed, and sometimes the date decides which engineering team is responsible. Reviewing thousands of signals by hand is often not feasible. The cost function, which defines "change", must be chosen up front, and often nobody can say which one fits or check the result.
Example from the article: a turbogenerator with more than 200 signals, where the goal was to find deviations from normal operation with an error of at most 1 day. A large drop in one signal was just a power reduction; the real anomaly was elsewhere.
Solution in brief
Treat retrospective diagnosis as offline changepoint detection: split the full series into segments by minimising the sum of a cost function over the segments, using the ruptures library. When the right cost function is unknown, scale several of different nature, aggregate them and run the usual search on the aggregate.
Questions this case answers
- How to locate the moments when a multivariate time series changes state when no labelled change points exist and the cost function that defines a change must be chosen up front
- How to choose a segmentation cost function when segmentation quality cannot be measured directly, by combining several cost functions into one aggregate
- How to find the moments when a multivariate time series changes state
- How to choose a cost function for change point detection when you cannot measure segmentation quality
Tools and methods
- ruptures — Python library for offline changepoint detection
- Cost functions — L1, L2, Mahalanobis-type, linear, autoregressive AR(1), kernel and others; each detects a different kind of change
- Search methods — Opt and Pelt (exact), Win, BinSeg, BotUp (approximate)
- Ensemble of cost functions — scaling (MinMax, Znorm, MinAbs, Rank) and aggregation (Min, Sum, WeightedSum, ThresholdSum)
- SKAB, Numenta Anomaly Benchmark (NAB) — datasets and metric of the examples
Google or a one-time email code. No password.