Blockwise Occupancy Inference under Weak Dependence with Record-Based Persistence Thresholds
We study the proportion of unusual diagnostic values within a time-series block. Two thresholds distinguish reference-compatible, elevated, and high occupancy. Under absolute regularity, a finite-sample count approximation accounts for the raw observations shared by successive diagnostics. A second calibration, based on reference-block ranks, accommodates fitted diagnostics and dependence within each block. We state the error contributions from fitting, block separation, and initialization, and calculate them for a nonlinear heteroskedastic model. Bounds on the diagnostic's response to a change give confidence sets for the fraction of affected positions; without such bounds, occupancy does not identify duration. We also characterize the effect of imposing a record-count threshold and the sensitivity of the count to replaced diagnostics. Simulations examine fitted calibration, reference length, calibration gaps, and power at comparable false-alarm rates. An application to bearing-vibration measurements quantifies the reduction in healthy high-occupancy labels and its sensitivity cost. The theoretical results give per-block guarantees and specify the additional assumptions needed to interpret high occupancy as persistence.
Authors
- Youssef Salman (ORCID: https://orcid.org/0009-0004-4401-5906)
- Hyam Abboud
- Anis Hoayek (ORCID: https://orcid.org/0000-0002-6499-6447)
- Mireille Batton‐Hubert
- Charbel Al Tannoury
Publication Details
- Journal
- HAL
- Published
- 2026-11-20
- Primary Topic
- Fault Detection and Control Systems
- Type
- preprint