Abstrak Artikel
This study proposes a rule-based Human-in-the-Loop (HITL) framework specifically designed for sensor-limited environments. The system integrates two data sources: a Numeric Health Score (NHS) calculated from five operating parameters, an Ordinal Utility Model (OUM) that quantifies qualitative assessments of field operators from routine inspection records. Both scores are fused through linear fusion with parameter α = 0.65 to produce a Health Index (HI) which is then classified into four Final Health Class (FHC) categories: HEALTHY (HI ≥ 75), WATCH (55 ≤ HI < 75), DEGRADED (35 ≤ HI < 55), and CRITICAL (HI < 35). Model uncertainty was validated through two Monte Carlo experiments (N = 10,000 iterations) using Dirichlet perturbation on AHP weights and Uniform perturbation on alpha fusion parameters, both yielding a Spearman rank correlation of ρ = 0.985. HI estimation uncertainty was quantified through bootstrap resampling (N = 10,000 per pump) resulting in a 95% Confidence Interval per pump. A three-signal based HITL gating mechanism (Confidence Level from bootstrap, Classification Stability Rate from Monte Carlo, and boundary proximity) automatically identifies pumps requiring further verification or review before FHC is assigned. The application of operational data on 16 pumps at TBBM (for 30 days or one month) resulted in: 11 HEALTHY pumps (68.75%), 4 WATCH pumps (25.0%), and 1 DEGRADED pump (6.25%), with 2 pumps requiring HITL review and 1 pump experiencing class downgrade through the HITL_DOWNGRADED path. This approach offers a condition-based maintenance solution that is auditable, reproducible, ready to be adopted in industrial facilities towards maintenance-digitization.
Pratinjau PDF
Pastikan PyMuPDF terpasang lalu coba muat ulang.