Beyond Detection Rate: What Transfers in Condition Monitoring—and What Does Not

This paper tests a simple, deterministic condition-monitoring pipeline for rotating and hydraulic machinery on five public data sets and reports where it works and where it does not. The pipeline is a banded envelope statistic over an order-tracked spectrum, a baseline frozen at commissioning, and a sequential evidence accumulator, with no learned model. The main point is that the transferable thing is not the detector but the commissioning record: which band was declared, which operating regimes were seen, when the baseline was frozen, and how good the tachometer is. Holding out whole bearings, it gets 0 false alarms on 480 held-out healthy records while naming extent-2 and extent-3 damage on 96 to 99 percent of records and most extent-1 damage almost never. The same pipeline under a band auto-selected from healthy data alone names nothing at all, so the band decision, made once at installation, is the difference between a working detector and a silent one. Order tracking is shown to be a precondition, with a tachometer scale-error cliff between 2 and 3 percent. Freezing the baseline is what makes slow degradation visible and is also what puts alarms on healthy hydraulic episodes; both sides are reported. Four things did not transfer: motor-current-only detection, a second variable-speed rig, a set of rotary actuators with no healthy history, and prognosis, where a pre-registered stopping rule fired and the claim was withdrawn. Every number is emitted from hashed result files that ship with the paper, and the text says which experiments did not run.