Calibrated Selective Disaggregation of Industrial Machinery Under Distribution Drift: A Multi-Week Real-World NILM Deployment

Replication package for the article "Calibrated Selective Disaggregation of Industrial Machinery Under Distribution Drift: A Multi-Week Real-World NILM Deployment" (submitted to IEEE Transactions on Industrial Informatics). Contents: 01_code (implementation, evaluation harness, five baseline implementations, figure generators, number-verification battery); 02_data (a derived, non-invertible per-window feature dataset plus every measured result file behind the reported numbers); 03_models (trained five-seed ensembles, released with the industrial partner's consent). A fail-closed battery re-derives all 133 reported numbers from the shipped artifacts, and a one-command check reproduces the released ensemble's behaviour on public data alone (pooled micro-MCC 0.857). The raw industrial measurements are not included and remain proprietary; the site is not identified, machines appear as M1-M7 and days as anonymised identifiers. The measurements are machine power signals and contain no personal data.