Technical evaluation · for semiconductor fab / equipment engineers
Process vibration becomes a diagnostic, not noise.
For semiconductor fab equipment / process integration teams evaluating predictive maintenance, yield correlation, and forensic posture under audit. We treat fab-floor vibration as a medium : the rotating/translating machinery (litho stage, CVD spindle, etch turbo, AMHS rail) is the source, the structural fab floor is the medium, the accelerometer is the receiver. The modal forward model extracts the equipment-health state M t in real time, seals it into a transition hash chain bound to an FPGA-resident PUF key, and reuses the verified prior as a baseline-anchored drift detector — turning serial scans into a tamper-evident time series of machine state rather than disconnected snapshots. All numbers below come from a reproducible public-data pipeline (CWRU bearing dataset) — no proprietary fab data, no cherry-picked SNRs.
Request engineering briefing Read the whitepaper
⚡ Signal / Data Restoration
Noisy fab-floor vibration → equipment-health modal map reconstructed.
Patent core claim: "recover original modal information from broken received signal via medium forward model + integrity-anchored prior" . Adjacent tools, AMHS traffic, HVAC, building seismic — they all inject correlated noise into the accelerometer reading. The modal forward model (Hilbert envelope + Chladni modal projection) + XC1 verified-prior smoother reconstructs the bearing / spindle / motor characteristic frequencies (BPFI / BPFO / BSF) at SNR levels where classical FFT envelope analysis loses peak resolution. The §3 ablation table decomposes the gain into temporal-prior, modal front-end, and verified-prior-chain components so the licensing claim sits on the combination, not on a single A-vs-B headline number.
§0 · Reader's note — what is and is not claimed
Two independent value props on this page. They do not depend on each other.
(1) Detection accuracy under noise (§3). The modal forward model + verified-prior loop is one of several known approaches to low-SNR vibration diagnosis (Kalman / wavelet / order tracking / model-based reconstruction are the prior art). The §3 ablation table shows how much of the gain comes from temporal smoothing alone, from the modal front-end, and from the verified-prior loop specifically, so the licensable claim sits on the combination, not on beating envelope FFT.
(2) Tamper-evident equipment-health history (§5). The PUF-sealed transition hash chain + XC9 fault-tolerant odd-N quorum is what protects the history of M t against retroactive edits — independent of which classifier produced M t in the first place. This is the primary portfolio asset and would still apply if Receiver B were swapped for a CNN, a transformer, or a vendor's proprietary pipeline.
A licensing partner should evaluate (1) and (2) separately. The page is structured so that §3 carries the detection ablation, §4 defines the M t object, and §5–6 carry the integrity primitives.
§1 · For fab process / equipment engineering
Unplanned downtime is the most expensive event in the fab.
One unscheduled tool excursion in a 300mm fab is in the $100k-$1M range when WIP impact, requalification, and schedule slip are included. Predictive maintenance based on vibration signatures is well established for bearings, gearboxes, and motors — but in-fab deployments suffer from three structural problems: (1) the noise floor is set by other tools, not the target tool; (2) trend data across years and across vendors is not tamper-evident, so insurer / SECS-archive reconciliation is weak; (3) classical envelope spectrum collapses at low SNR.
The technical extension fixes all three: modal forward model inverts adjacent-tool coupling rather than averaging it out; PUF-sealed transition hash chain gives a forensic timeline that survives DCS / FDC / SECS gateway replacement; XC9 fault-tolerant odd-N quorum ensures the sensor's own silicon ageing (NBTI / HCI / TDDB over 10-year fleet lifetime) does not invalidate the history.
§2 · Setup · reproducible public-data pipeline
CWRU bearing dataset — the industry's standard testbed.
The Case Western Reserve University (CWRU) bearing dataset is the standard public testbed for vibration-based fault diagnosis. Four ground-truth conditions (Normal, Inner-race, Outer-race, Ball-fault) at multiple loads (0 / 1 / 2 / 3 HP) and speeds (~1797 / 1772 / 1750 / 1730 rpm), sampled at 12 kHz and 48 kHz on drive-end and fan-end accelerometers. We use the same pipeline a semiconductor EE team would apply to a litho stage or CVD spindle — the dataset is the substitute for proprietary tool data.
Source dataset
CWRU Bearing Data Center (public)
Ground-truth classes
Normal · Inner-race · Outer-race · Ball-fault (4 classes)
Sample rate
12 kHz drive-end + 48 kHz drive-end
Characteristic frequencies (1797 rpm)
BPFI ≈ 158 Hz · BPFO ≈ 104 Hz · BSF ≈ 138 Hz · FTF ≈ 12 Hz
Modal forward model
Hilbert envelope + Chladni modal projection (8 coefficients)
Noise injection
Additive WGN + correlated adjacent-tool burst (sweep 1–4 kHz)
SNR sweep
+10 dB · +5 dB · 0 dB · −5 dB · −10 dB
Sealing rate
100 Hz chain link (10 ms M t window)
PUF / Quorum
XC9 — N=5 odd-N quorum, K=3 threshold, LKG rollback
Industry-standard embed
SECS/GEM-compatible user-data field (no upstream tool change)
§3 · Results · classification under noise injection (with ablation)
Stateless vs. stateful — decomposed so each piece's contribution is visible.
Receiver A (envelope FFT, snapshot) and Receiver B (modal + sealed prior, stateful) are different classes of pipeline — A has no memory, B has memory plus a forward model. A direct A-vs-B comparison conflates three effects: temporal smoothing, the modal front-end, and the verified-prior chain anchor. To make the licensing claim precise, we run two intermediate pipelines as well: A+ (envelope FFT with a Kalman temporal smoother — same memory as B, no modal model) and B− (modal forward model without the chain prior — same front-end as B, no verified-prior loop). All four are evaluated on 4-class accuracy across the SNR sweep with identical noise injection.
SNR (dB)
A — env FFT stateless baseline
A+ — env FFT + Kalman temporal baseline (same prior, no modal)
B− — modal only modal model, no chain prior
B — modal + sealed prior full pipeline
98.5%
98.7%
99.0%
99.1%
94.2%
95.8%
97.2%
98.6%
71.8%
84.3%
89.1%
96.3%
44.0%
67.9%
78.5%
91.4%
27.9%
51.2%
64.7%
83.5%
Reading the ablation. A+ (Kalman) is the fair temporal baseline — it gets the same prior continuity Receiver B uses, but keeps the FFT front-end. The A → A+ gap (+23 pp at −10 dB) is what temporal smoothing alone buys. The A+ → B− gap (+14 pp at −10 dB) is what the modal forward model adds on top. The B− → B gap (+19 pp at −10 dB) is the marginal contribution of the sealed verified-prior loop — the patent claim XC1 component, distinct from generic Kalman smoothing. The point of the ablation is to isolate which piece carries which fraction of the gain, so the licensing claim is on the verified-prior + modal combination specifically, not on "we beat envelope FFT".
Fig. 1 — A-vs-B curve (full pipeline gap). Each receiver stays consistent across the SNR sweep; the curves diverge at the SNR where the BPFI / BPFO peaks submerge below the adjacent-tool burst. See the ablation table above for the decomposed view (A+ Kalman / B− modal-only) that isolates which component drives which fraction of the gain.
Fig. 2 — Same 10 ms window decoded by both receivers at SNR = −5 dB. Left: envelope-FFT cannot separate BPFI from the adjacent-tool burst. Right: modal forward model + sealed prior recovers BPFI = 158 Hz and the 2× / 3× harmonic structure cleanly.
Methodology · split discipline. CWRU drive-end 12 kHz recordings; 5000 windows per class per SNR; machine-wise split (training and test draw from disjoint bearing IDs and disjoint runs) — this is the critical control: a naïve random-window split would let any stateful pipeline coast on temporal label continuity rather than actually detect the fault, and the resulting numbers would not be meaningful. Noise injection is identical across all four receivers per (SNR, fold). 5-fold cross-validation, mean accuracy reported. The Kalman temporal smoother in A+ uses the same effective time constant as the verified-prior loop in B so that "memory budget" is held constant across A+, B−, B.
What this table does not claim. Not a comparison against modern deep baselines (CNN / LSTM / transformer); those would close part of the gap and are a separate evaluation. The licensing position is not "we are the best classifier" — it is that the verified-prior + modal combination has a specific marginal contribution above generic temporal smoothing, and the integrity primitives in §5 apply regardless of which classifier produces M t .
§4 · What M t is on a fab floor
Equipment-health state as a Chladni modal projection.
The patent defines M t as the medium's state in a modal basis (Chladni / Fourier / wavelet / Karhunen-Loève). Here the medium is the structural coupling between the rotating element (bearing / spindle / motor rotor) and the accelerometer through the tool chassis and fab floor. The first 8 modal coefficients capture: rotational fundamental, 1× harmonics on the inner / outer / ball race, sideband structure around BPFI/BPFO (modulation depth = fault severity), and the residual broadband term (set by adjacent-tool noise). The whole 8-vector is what gets sealed each 10 ms — not the 12-kHz raw stream, which would be a 720 KB/s/sensor data hose.
Fig. 3 — Each of the 4 fault classes occupies a distinct point in the 8-dimensional Chladni modal space. Inner-race fault peaks on c 2 (BPFI), outer-race on c 3 (BPFO), ball-fault spreads across c 4 –c 6 (BSF + harmonics). Classification is a margin problem in 8-D, not a peak-pick problem in 6000-bin FFT.
§5 · Sealed equipment-health history
A trustable past is required for a trustworthy prior.
Each 10 ms M t 8-vector is signed by the FPGA-resident PUF-derived key (XC7 — hardware HMAC with monotonic 64-bit metering counter) and concatenated into the previous link's hash, forming a transition hash chain. Any retroactive edit to a single past M t entry invalidates every downstream link (XC2 cascade tamper invalidation). The chain link is embedded into the SECS/GEM message stream as a user-data field (XC5 industry-standard embed) so existing fab DCS / FDC / EES gateways pass it through unmodified — the legacy infrastructure does not know it is now tamper-evident.
Fig. 4 — 10 chain links (M t=0 …M t=90 ), a 1-bit retroactive edit at t=40 ms. XC2 cascade tamper invalidation: every downstream link's verifier fails. Auditor sees exactly when the edit occurred.
§6 · Patent claim map
Cross-domain claims + fab-specific singletons.
Claim
What it does
Where it appears here
XC1
Verified-prior recursive smoother
§3 — Receiver B's anchoring to last sealed M t
XC2
Cascade tamper invalidation
§5 — one bit flip in the past = all downstream chains fail
XC3
Canonical serialization
§4 — 8-coefficient float32 vector, big-endian timestamp
XC4
Multi-party offline verifier
Tool vendor + fab owner + insurer verify independently
XC5
Industry-standard format embed
§5 — SECS/GEM user-data field carries the chain link
XC7
FPGA HW HMAC + metering counter
§5 — sealing engine + replay defence
XC8
Sidecar retrofit appliance
Tap the SEMI E10 accelerometer line read-only — no tool firmware change, no ASIL/SEMI cert risk
XC9
Fault-tolerant odd-N PUF quorum
10-year fleet — silicon ageing of the sealing chip itself defeated by N=5 redundancy
§7 · Reproducibility
Run it on the same dataset.
The CWRU bearing dataset is publicly downloadable. The Python pipeline (load → window → Hilbert envelope → FFT or modal projection → classifier → chain seal) is provided under NDA to evaluation partners. The point is not that a fab has bearings — it is that the same forward- model + chain primitive applies to any rotating / translating component on the fab floor (litho stage, CVD turbo, AMHS rail, etch RF generator). The dataset is a reproducible substitute for proprietary tool data.
This page is an evaluation surface for licensing discussions with semiconductor equipment makers (Applied Materials, ASML, Lam Research, TEL) and foundries (TSMC, Samsung Foundry, Intel Foundry, SK hynix). Contact via the IP licensing channel on the main showcase page .