Stratum — Industrial Anomaly Detection Demo

Fault Wave Progress

vibration RMS kurtosis RPM temperature motor current

Inner Race

F1 0.911

PENDING

Outer Race

F1 0.819

PENDING

Ball Fault

F1 0.758

PENDING

Signal Legend

Signal A — PSV cosine distance > θ_a (structural deviation)

Signal B — Fault token prevalence > θ_b

Signal C — Kurtosis > 95th percentile (early bearing fault)

Calibration Profile

Status—

n_samples—

theta_a (Signal A threshold)—

theta_b (Signal B threshold)—

kurtosis_threshold (Signal C)—

Calibration latency—

Fault data usedNONE

CWRU Bearing Benchmark

Fault Type	F1 Score
Inner Race	0.911
Outer Race	0.819
Ball Fault	0.758
Overall	0.832

How Stratum works: Stratum calibrates on 80 healthy sensor windows — vibration RMS, kurtosis, RPM, temperature, and motor current. No fault samples, no labels, no model retraining required. The PCF engine builds a structural fingerprint of normal machine behavior using Positional Correlation Fields. When a bearing fault develops, three signals fire independently: Signal A detects structural deviation in the PSV cosine distance; Signal B detects fault token prevalence exceeding the calibrated threshold; Signal C detects kurtosis exceeding the 95th percentile — the earliest indicator of bearing fatigue. A fault verdict requires only one signal to fire (three-signal OR).

CWRU Bearing Dataset — Inner Race faults are the strongest signal (F1 0.911) due to sharp kurtosis spikes. Ball faults are the most subtle (F1 0.758) and rely primarily on Signal A structural deviation. Patent-pending algorithm. Patent U.S. Provisional No. 63/978,633.