External transfer test. Freeze before calculating features or model scores on Hanoi data. The source article, file names, shapes, shaft-frequency metadata and bearing geometry were inspected for design. HUST did not supply training observations, tuning or alarm cutoffs. Scope is a specified bearing embodiment of US20260220223A1, not authentication of historical production code or validation of every optional patent branch.
Training: first accelerometer, first 2 seconds of all 96 recordings from the first VesselSeed study (Saarland, Zenodo 15376390). Ridge logistic scores fitted after training-only standardization. Comparator is speed-adjusted envelope power at first three inner-ring harmonics from the same band and window. Both arms see the same channel, time, speed and geometry. Test: all 15 normal and 15 isolated inner-race HUST v3 steady-speed files across 6204–6208 and 0/200/400 W. One observation per recording; do not split into windows. HUST fs is shaft Hz; original sampling 51.2 kHz; resample to 20 kHz. BPFI from published dimensions, pitch approximated by average of bore and outer diameters and zero contact angle. Ignore run-up, mixed and other fault types. Record SHA-256 per MAT file.
Primary: false positives at an external retrospective target of at least 90% fault detection. With 15 positive cases, select rank 14; report actual counts in case of ties. Also report operating performance using thresholds learned on Saarland with no adjustment on HUST. No re-fitting after results. Report per-recording scores, no-peak and nonconvergence counts. Repeated loads on a rig are correlated. No independent field or within-record speed-ramp claim.
Method map: direct waveform input; bandpass, envelope and Fourier decomposition; anchor frequencies scaled as A ln(p) with A set from shaft speed and geometry; measured conjugate Fourier angle rotates each constituent to its cosine reference at the window origin. 7x7 counterclockwise spiral, first 15 primes, alternating handedness, nearest band peak, mismatch decay, harmonic fit and coherence from equation 2. Fill vacant regions from populated neighbors (table 3). Four to 32 synchronous equation-7 updates with alpha .015 on normalized local chirality/geometry weights; use maximum neighborhood coherence discrepancy <= .90. Score eligible prime anchors with local multiharmonic alignment (.6/.3/.1), coherence, compatibility and positive observed log power ratio; select strongest per hand with L and R phase filters using wrapped angle magnitude. Missing hand contributes zero. No converged field gives zero grid evidence. Frozen bearing-specific readout of the two peaks; also report grid alone, primary baseline plus full grid and same fitted combined model with observed offsets set to zero before filling the grid. Fixed spatial form and input/output adapter; optional subanchors, wavelet filters and memory not used. Full training and test code is supplied.
