Stratum applies Positional Correlation Fields to industrial systems — detecting equipment faults from sensor data with no training labels required, and validating data center designs against certified reference patterns before a single server is racked.
F1=0.832
CWRU Bearing Benchmark
~1.3ms
Detection Latency
CPU
No GPU Required
0
Fault Labels Needed
⚙ Industrial Anomaly Detection
▦ Data Center Design
Bearing Fault Detection
Three Fault Types. No Training Data.
Stratum calibrates on 80 healthy sensor windows — vibration RMS, kurtosis, RPM, temperature, motor current. No fault samples, no labels. When a bearing fault develops, structural deviation from the healthy baseline fires one or more signals independently.
⚙
Inner Race Fault
Strongest Signal — All Three Fire
Sharp kurtosis spikes from periodic ball-race impact at the characteristic defect frequency. Signal A (PSV structural deviation), Signal B (fault token prevalence), and Signal C (kurtosis 95th pct) all fire independently.
0.911
F1 Score
0.891
Precision
0.932
Recall
◎
Outer Race Fault
Moderate Kurtosis + Structural Shift
Sustained vibration elevation with moderate kurtosis. Signal A (PSV cosine distance) and Signal C (kurtosis threshold) fire; Signal B near boundary. Consistent thermal elevation from friction at outer race contact.
0.819
F1 Score
0.801
Precision
0.838
Recall
●
Ball Element Fault
Subtle — Signal A Only
The hardest fault class. Ball defects produce subtle load variation with minor structural deviation. Only Signal A (PSV cosine distance) fires reliably — kurtosis stays near-normal, token prevalence is boundary-level.
0.758
F1 Score
0.741
Precision
0.776
Recall
Three-Signal OR Classifier
Any Signal Fires → Fault Detected
A fault verdict requires only one of three independent signals. This means Stratum catches subtle faults that individual methods miss — while keeping false positive rates low since each signal has its own calibrated threshold.
A
PSV Structural Deviation
Cosine distance between the current window's Positional Scoring Vector and the calibrated normal centroid. Catches broad structural change in how sensor values co-occur — fires for all fault types.
Fires for
Inner RaceOuter RaceBall Fault
B
Fault Token Prevalence
Fraction of fault-indicating tokens (VIB_ZONE_C, VIB_KURTOSIS_HIGH, TEMP_ELEVATED) in the tokenized sensor window exceeds the 99th percentile threshold from calibration. High specificity for severe faults.
Fires for
Inner Racesevere only
C
Kurtosis 95th Percentile
Vibration kurtosis exceeds the 95th percentile derived from calibration — the earliest indicator of bearing fatigue. Kurtosis elevation precedes RMS elevation in bearing faults, giving Signal C early-detection advantage.
Fires for
Inner RaceOuter Race
CWRU Bearing Benchmark
Validated on Real Bearing Data
Case Western Reserve University bearing dataset — the standard benchmark for bearing fault detection. 2,800+ windows across three fault types at multiple fault diameters and load conditions. No fault data used in calibration.
PCF tokenizes each design layer independently and computes a PSV — then scores the structural coherence of the design as a whole against reference patterns from certified, high-efficiency builds.
Layer 01 — Physical
Rack Layout & Aisle Design
Hot-aisle/cold-aisle containment, rack density distribution, inter-row spacing, and perimeter clearances are tokenized as a structural sequence and scored against ASHRAE A1/A2 reference layouts.
HOT_AISLECOLD_AISLERACK_DENSEAISLE_GAPCONTAINMENT
Layer 02 — Thermal
Cooling Architecture
CRAC/CRAH unit placement, CFM ratings relative to rack load, supply and return air path topology — scored against TIER-III baseline thermal patterns for adequate redundancy and airflow coverage.
UPS sizing, PDU placement, circuit redundancy, generator capacity, and feed diversity — tokenized into a power topology sequence and compared against Uptime Institute TIER ratings for structural completeness.
Spine-leaf vs. three-tier topology, uplink over-subscription ratios, east-west bandwidth provisioning, and cross-connect diversity — scored against reference patterns for hyperscale and colocation designs.
Each design parameter — rack count, cooling redundancy level, power feed configuration — maps to a semantic token in the PCF vocabulary. The design becomes a structured token stream per layer.
2N UPS → PDU_REDUND
Step 02
Compute Design PSV
Positional Scoring Vectors are computed from the token stream — capturing how design choices co-occur structurally. A dense rack zone adjacent to undersized cooling produces a distinctive structural signature.
proprietary structural scoring
Step 03
Compare to Reference
The design's PSV is compared against reference PSVs from TIER-III/IV certified builds and ASHRAE best-practice templates. Structural distance is computed per design layer independently.
structural divergence score
Step 04
Score & Flag
Each layer receives a structural score. Deviations above threshold are flagged with the specific design co-occurrence that caused them — actionable before construction begins.
divergence above threshold → FLAG
Design Analyzer
PCF Design Scoring
REFERENCE: TIER-III BASELINE
Rack Layout
Cooling Redundancy
Power Topology
Network Fabric
PSV vs. Tier-III Reference
Design Findings
Target Applications
Colocation & Hyperscale
Score greenfield designs before construction. Validate expansions against existing certified PSV baselines. Reduce costly design-phase rework.
Design/Build Firms
Give clients structural validation scores alongside traditional CFD simulation — faster, cheaper first-pass design review without a full simulation cycle.
Enterprise IT Architects
Validate on-premises data center expansions against ASHRAE and Uptime Institute reference patterns. Catch single points of failure before they become incidents.