Hanoi motor startup | frozen comparison, 21 November 2026

Question: On real accelerating recordings, at equal fault detection, does retaining measured phase save any healthy alarms versus both conventional order tracking and the same fitted model with measured phase removed? HUST Mendeley v3, DOI 10.17632/cbv7jyx4p9.3. Previously inspected steady-state data, archive headers, some RPM endpoints and the authors' descriptions. No startup vibration scores inspected before freeze. Same rig and potentially the same bearings as the previous Hanoi study: a new condition, not an independent replication.

Eligibility: inventory all 30 normal (N) and isolated inner-race (I) files in the five published sizes and three loads. Other fault classes are outside the fitted detector. Use ru vibration and synchronous rpm, 51,200 Hz, equal length. Never substitute the steady-state data/fs, unpaired ru_raw, estimated speed, or a reversed slowdown. Disclose every excluded recording. Require one finite two-second segment with all RPM 60..2400, at least 200 RPM increase between median first/last 100 ms, and >=80% of consecutive 100 ms bin medians rising or level. At 100 ms start increments choose greatest qualifying gain, ties earliest. Selection sees RPM only. One decision per file, exact duplicate waveform/shaft trace pairs count once, conflicting duplicate labels abort. No retuning. Report size/load strata represented in both classes separately to expose confounding.

Processing: use the same two seconds for all arms. Anti-alias vibration to 20 kHz, apply the previous 1–7 kHz filter and Hilbert envelope in time. Integrate measured shaft speed to accumulated turns and interpolate the envelope to 40,000 equal-angle observations; use observed mean shaft frequency to map spectral orders onto the previous feature scale. Same published geometry and inner-race order for both arms. Reuse the prior full grid, model coefficients and standardization fitted only on Saarland. Baseline = three inner-race envelope harmonics. Phase on = baseline plus full grid. Phase off = same inputs and combined weights with measured grid phase offsets zeroed. No standalone grid claim.

Primary comparison: on all eligible files, retrospectively alarm at the highest threshold detecting at least 90% of the labelled defects in each arm (inclusive ties; ceiling rank). Report exact true/false positives and identifiers of discordant calls. Separately use the existing Saarland thresholds unchanged for practical decisions. Also report counts on eligible common size/load strata with thresholds recomputed at matched detection within that subset and the same locked thresholds. If either class absent at report level, report matched performance unavailable. Fewer false alarms than both controls at the same actual detection count is a descriptive win only; a tie or loss is published. Report all denominators and missing inputs. Distinct recordings and loads do not establish independent machines, predictive value for a plant, or field savings.
