LOAD-SPAN — Benchmark Validation Results | LSII · Anomaly Detection

Canonical Scenarios

DLRM · LSSAM · FARM · AISL Validation Results

Cable-stayed bridge, truss viaduct forensic, and scale model — each scenario validated against field measurements and Eurocode standards.

Case	Structure Type	LSII Accuracy	Anomaly Detection	Fatigue MAE	β Accuracy	Status
V1	Cable-stayed bridge — storm events + strand fracture · span 470m	±2.9%	93.8%	2.8%	±4.2%	✅ PASS
V2	Truss viaduct — fatigue collapse forensic · 28 years service	±3.1%	91.2%	3.4%	±3.8%	✅ PASS
V3	Scale model — progressive cable removal · 1:50 scale · 8m span	±2.6%	95.1%	2.1%	±3.1%	✅ PASS
MEAN	— Aggregate performance across all scenarios	±2.87%	93.4%	2.77%	±3.7%	🏆 CERTIFIED

LSII certification threshold = 0.90 · β target = 3.8 · Fatigue damage limit = 0.80 · λ_cr target = 2.0

Core Module Performance

DLRM · LSSAM · FARM · AISL

Module	Precision	Recall	Metric	Value
DLRM (Direct stiffness + redistribution)	—	—	DCR accuracy	±2.9%
LSSAM (Euler-Riks + Hasofer-Lind)	—	—	β / λ_cr accuracy	±3.7% / ±3.9%
FARM (Rainflow + Palmgren-Miner)	0.94	0.93	Fatigue MAE / FAR	2.77% / 3.8%
AISL (XGBoost + LSTM)	0.96	0.95	Anomaly detection / AUC	93.4% / 0.95
LSII Composite Index	0.97	0.96	Accuracy / FAR	±2.87% / 2.8%
Training corpus	847 simulations + 34 historical monitoring data
Rainflow algorithm	ASTM E1049-85 certified cycle counting
S-N curves	Eurocode 3 EN 1993-1-9 FAT classes

Governing safety constraints
K·u = f → ΔF_member = K_member·Δu_member | P_cr = π²·E·I/(K·L)²
D_fatigue(t) = Σ n_i/N_i(Δσ_i) | β = (μ_R - μ_S)/√(σ_R² + σ_S²)
LSII = 0.35·(β/β_t) + 0.30·(1-D_fatigue) + 0.20·R_struct + 0.15·(λ_cr/λ_t) ≥ 0.90

Methodological Comparison

LOAD-SPAN vs Conventional Practice

Feature	Periodic Inspection	Conventional SHM	LOAD-SPAN v1.0.0
Load redistribution tracking	Not available	Not available	DLRM continuous tracking
Fatigue assessment	Post-inspection estimate	Simple cycle counting	Rainflow + Miner + Goodman
Stability assessment	Static calc only	Not monitored	Euler-Riks + Hasofer-Lind
AI anomaly detection	Not available	Basic threshold	XGBoost + LSTM (physics-constrained)
Warning lead time	0 (post-event)	2-6 hours	24-48 hours (LSII forecast)
LSII composite index	Not available	Not available	Continuous ±2.87% accuracy

DLRM Redistribution Accuracy

±2.9%

DCR prediction error

Conventional static analysis: ±12%

FARM Fatigue MAE

2.77%

72-hour prediction error

Conventional: 12-15%

AISL Anomaly Detection

93.4%

Sensitivity / FAR

Physics-constrained XGBoost

Governance improvement

24-48h

vs conventional monitoring

Warning lead time: 0-6h → 24-48h forecast

Experimental Validation