The Tier-0 Condition-Monitoring Gate: An Out-of-Sample Test of a Frozen Single-Token Encoder in Industrial Predictive Maintenance

Abstract

The Tier-0 Condition-Monitoring Gate: An Out-of-Sample Test of a Frozen Single-Token Encoder in Industrial Predictive Maintenance Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: 10.5281/zenodo.23127675 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign · a characterization of previously filed and published methods; no new algorithmic subject matter is disclosed. Abstract A companion capstone proposed a predictive theory of a frozen, deterministic, class-discriminant single-token encoder used as an always-on Tier-0 sensing layer, stated as four falsifiable laws — a competence zone (L1), per-entity necessity (L2), compose-versus-re-commission (L3), and cost-not-accuracy (L4) — and validated it out-of-sample on a chemical e-nose and, in follow-ons, on radio-frequency modulation classification and device fingerprinting. Those tests exercised the laws on sensing and identity tasks but not on the application the encoder was built for: a battery- and radio-constrained edge node that must watch a machine for years and wake an expensive downstream path only when a fault is plausible. We supply that test here, on industrial predictive maintenance, as a rigorous out-of-sample deployment characterization. Across three independent bearing-vibration corpora (a cross-load benchmark, a record-disjoint benchmark, and a hard bearing-disjoint real-accelerated-damage corpus of 2,880 windows over nine bearings and four operating conditions) and four distinct machine-acoustic types, with a medical electrocardiogram (ECG) gate as an out-of-zone control, six pre-registered predictions hold. L1: the token is at near-parity with a strong RandomForest on vibration fault detection — and on the hard real-damage corpus it slightly edges the baseline (AUROC 0.909 vs 0.897, bearing-disjoint) — while paying a real, codebook-robust accuracy tax on ECG morphology (+0.05 to +0.19), where the decision is entangled with individual identity. L2: machine wear is entity-specific — each machine's own label-free "self-twin" detects its faults (own ~0.94–0.96) while another machine's normal model is near chance (cross ~0.42–0.54), and a supervised cross-machine gate fails for the token and the baseline alike; industrial wear thereby joins the biological and hardware poles as per-entity-necessary. L3: the normal model does not transfer across the operating envelope — a benign speed change reads as a fault (up to ~96–100% false alarms) unless the model is re-commissioned or condition-aware thresholds are used, a category limit shared by every one-class detector, while the supervised fault classifier does transfer across loads. L4: on raw accuracy the token is not supreme — a Gaussian mixture ties it on acoustic and a RandomForest beats it under injected sensor noise — so the defensible edge is the unified, tiny (≤256-cell), bit-exact, auditable artifact: a single frozen codebook that yields both the classification decision and the label-free monitor, at near-parity, inside a two-tier escalation that at realistic fault prevalence runs at 3–11% of an always-on full model's compute on vibration. A negative control closes the loop: appending a three-bit "regime" code derived from the same embedding does not beat a plain token or a generic code — the single-token accuracy tax is walled, consistent with L1. We disclose the two real weaknesses (noise fragility; operating-condition-shift false alarms), their partial mitigations, and the honest conclusion that the value here is cost, determinism, and auditability — not accuracy. This is the third out-of-sample domain in which the four laws have held without amendment. No new algorithmic subject matter is disclosed; every result is a property of a previously filed and published method. Highlights · Competence zone (L1) holds on vibration — the token is at near-parity with a strong RandomForest, and on a hard, bearing-disjoint, real-accelerated-damage corpus (Paderborn) it slightly edges the baseline (0.909 vs 0.897), rebutting the concern that parity rests on saturated benchmarks; it pays a real, codebook-robust tax only on ECG morphology (+0.05 to +0.19), where the decision is entangled with individual identity. · Per-entity necessity (L2) confirmed for machine wear — each machine's own label-free self-twin detects its faults (own ~0.94–0.96) while another machine's model is near chance (cross ~0.42–0.54); supervised cross-machine transfer fails for the token and the baseline alike. Industrial wear joins the biological and hardware poles as per-entity-necessary, opposite the universal-extreme pole. · Compose vs re-commission (L3), with the supervised/unsupervised scope made explicit — the unsupervised healthy-only monitor must be re-commissioned across the operating envelope (a benign speed change triggers up to ~96–100% false alarms unless condition-aware thresholds are used, a category limit of all one-class detectors), while the supervised fault classifier transfers across loads unchanged. · Cost not accuracy (L4) — the token is not accuracy-supreme (a Gaussian mixture ties it on acoustic; a RandomForest beats it under injected noise), and a Gaussian mixture is in the same tiny footprint class; its defensible edge is the unified, tiny (≤256-cell), bit-exact, auditable artifact — one codebook yields both the decision and the label-free monitor — inside a two-tier escalation at 3–11% of an always-on full model's compute on vibration. · A grounded energy / telemetry ledger sharpens the economics and corrects an earlier over-optimistic figure of our own: transformative versus streaming raw windows (~150–2,500×), roughly neutral versus a competent on-device model, because the dominant feature-extraction cost is shared by both tiers. · Cheap to commission, broad in scope, and deterministic — the label-free self-twin stands up on a new machine in a handful of healthy windows on vibration (~0.99 by 20 windows on MFPT; 0.96 by 20 on Paderborn), more slowly on acoustic; the result spans three independent bearing-vibration corpora, four machine-acoustic types, and an ECG out-of-zone control, and the token is bit-exact (32- vs 64-bit agreement 1.000) with an auditable ≤256-entry lookup table. · A pre-registered negative control (the regime "backpack") confirms the single-token tax is walled — re-processing the token with information already inside it does not help — and a sweep of the platform's customization dials does not convert near-parity into superiority. What this record contains · Manuscript_Paper56.pdf — the industrial-PdM four-laws test (6 figures, 36 references); and Manuscript_Paper56.docx, the editable source. · PAPER_56_ZENODO_ARCHIVE.zip — the reproducibility archive (md5 in ARCHIVE_MD5.txt): the two frozen pre-registrations (PDM_GATE_PREREG.md, ICM_GATE_PREREG.md); the industrial gate and vibration-breadth runners, the ECG out-of-zone control, the self-twin-versus-unsupervised-panel comparison, the robustness / stress-test runners, the hard real-damage-corpus runner, the platform-option sweeps, the regime-backpack negative control, and the grounded energy / telemetry ledger builder; the frozen token encoder module; the result records (JSON); the six figures; the figure builder; the manuscript source; and a README. Public data is not redistributed (CWRU, MFPT, Paderborn KAt, MIMII, and AF-DB / MIT-BIH / PTB-XL, fetched or read from a local cache at run time); all paths and handles are scrubbed (PATH_TO_DATA / PATH_TO_SCRATCH, any cloud handle → MODAL_USER) and leak-scanned, and the per-deployment configuration-selection procedure is retained privately and excluded. Cite as R. J. Ferlic and K. K. Ferlic, "The Tier-0 condition-monitoring gate: an out-of-sample test of a frozen single-token encoder in industrial predictive maintenance," Zenodo, 2026, doi: 10.5281/zenodo.23127675. License and patent notice Released under CC-BY 4.0. Consistent with that license, no patent or other intellectual-property right of the authors is licensed, waived, or conveyed by this deposit. This work characterizes previously described, filed, or published methods and discloses no new algorithmic subject matter; random forests, gradient boosting, k-means / vector quantization, Fisher discriminant analysis, the information bottleneck, nearest-centroid / Mahalanobis / isolation-forest / one-class-SVM / Gaussian-mixture / reconstruction novelty detection, and cascade / early-exit inference are established prior art, used only as tools. The methods characterized — the class-discriminant single-token codebook encoder and its nearest-centroid monitor, the per-entity self-twin, the multi-token / token-ladder and soft-readout mechanisms, and the inference-time co-channel fusion — are the subject of filed and pending U.S. patent applications held by the authors, spanning the encoder, the personalization / on-device-adaptation applications (which include the per-entity self-twin), the multi-token / token-ladder application, and the inference-time co-channel fusion application. The competence-zone, per-entity-necessity, re-commission, cost / determinism, and two-tier-economics results reported here are properties of the frozen filed method, not new subject matter; the two-tier escalation is established cascade / early-exit inference combined with the filed gate. The per-deployment configuration-selection and commissioning procedure is retained as a trade secret and is not disclosed here. The capability is characterized for machine-health monitoring; no company or product is named, and the framing is generic to a battery-powered wireless condition-monitoring node and an expensive downstream path. © 2026 Fieldstone Analytics, LLC and the authors. Inquiries: randolphf@fieldstoneanalyticsllc.com. Companion deposits (spiral-domain-encoder-campaign) · The Frozen Token — the predictive-theory