Unsupervised Offline Changepoint Detection Ensembles

Offline changepoint detection (CPD) algorithms are used for signal segmentation in an optimal way. Generally, these based on the assumption that signal’s changed statistical properties known, and appropriate models (metrics, cost functions) used. Otherwise, process of proper model selection can become laborious time-consuming with uncertain results. Although ensemble approach is well known increasing robustness individual dealing mentioned challenges, it weakly formalized much less highlighted CPD problems than outlier or classification problems. This paper proposes unsupervised (CPDE) procedure pseudocode particular proposed link to their Python realization. The approach’s novelty aggregating several functions before search running during offline analysis. numerical experiment showed CPDE outperforms non-ensemble procedures. Additionally, we focused analyzing common algorithms, scaling, aggregation functions, comparing them experiment. results were obtained two anomaly benchmarks contain industrial faults failures—Tennessee Eastman Process (TEP) Skoltech Anomaly Benchmark (SKAB). One possible applications our research estimation failure time fault identification isolation technical diagnostics.