SMARTPHONE CHASSIS VIBRATION RESPONSE PROFILING A ONE-CLASS ENSEMBLE METHOD WITH DISTRIBUTION-FREE FALSE-ALARM GUARANTEES

We present a novel anomaly-detection method for smartphone chassis vibration response profiling that
combines four complementary one-class detectors with split-conformal calibration to deliver distribution-free
false-positive rate guarantees. The primary methodological contribution is a cross-modal coupling head that
exploits a structural property the per-channel detectors miss: optical camera-derived sub-pixel motion and
inertial accelerometer sensors are two distinct sensor views of the same chassis vibration, linked by the physical
transfer function. A healthy chassis preserves the optical-inertial linkage; mechanical defects perturb it even
when neither channel individually leaves its envelope. We formalise the coupling head as the symmetric
prediction MSE of two ridge regressors fitted on healthy data and show that it adds non-redundant
discriminative signal over per-channel methods. The four heads are combined by rank-average against per-head
healthy calibration distributions, then thresholded by a split-conformal quantile that provides a finite-sample
false-positive-rate bound under exchangeability. On the held-out evaluation split (100 healthy fingerprints and
88 verified-damaged units in four defect classes, drawn from a 712-fingerprint corpus across 356 physical
devices spanning eight flagship Android models), the method achieves true-positive rate TPR = 0.909 (80/88
damaged units detected), false-positive rate FPR = 0.030 (3/100) at the conformally-calibrated α = 0.05
operating point, and ROC AUC = 0.954 on the binary healthy-vs-damaged task. Against the production
diagonal-Mahalanobis baseline this corresponds to a 12× reduction in FPR (36/100 to 3/100) on the held-out
healthy split. Separately, in deployment, the system has processed 15,000 production scans across 4,800 devices
over an eight-month window, with 99.8% first-pass cryptographic verification success; on the same input stream
the legacy baseline issued 25.3× more anomaly verdicts and 6.8× more total alerts than the ML ensemble.
Production alert-rate differences are not the same quantity as evaluation-corpus FPR reduction and we report
them separately throughout. The on-device fingerprint pipeline runs in pure Kotlin with no on-device ML; the
ensemble runs server-side via ONNX Runtime.