MAD System Validation Results
Validation results for VLTHRLAB's LSTM autoencoder anomaly detection system across 8 public benchmark datasets.
Per-Dataset Validation Results
| Dataset | AUC | Detection Window (days) | False Alarms |
|---|
| Dataset 1 (Bearing) | 1.000 | 15 | 0 |
| Dataset 2 (Bearing) | 0.9998 | 12 | 0 |
| Dataset 3 (Gearbox) | 1.000 | 18 | 0 |
| Dataset 4 (Hydraulic) | 0.9999 | 8 | 0 |
| Dataset 5 (Valve) | 0.9997 | 10 | 0 |
| Dataset 6 (Compressor) | 1.000 | 24 | 0 |
| Dataset 7 (Generator) | 0.9995 | 5 | 0 |
| Dataset 8 (Pump) | 0.9999 | 14 | 0 |
Cross-Dataset Transfer Results
| Train Dataset | Test Dataset | AUC | False Alarms |
|---|
| Dataset 1 (Bearing) | Dataset 6 (Compressor) | 0.9995 | 0 |
| Dataset 3 (Gearbox) | Dataset 8 (Pump) | 0.9993 | 0 |
| Dataset 1 (Bearing) | Dataset 3 (Gearbox) | 0.9997 | 0 |
Digital Twin Simulation Results
| Metric | Run-to-Failure | Predictive Maintenance | Improvement |
|---|
| Downtime per event | 40 hours | 11.2 hours | 72% reduction |
| Cost per event | $1,000,000 | $144,000 | 86% reduction |
| Maintenance cost ratio | 6x baseline | 1x baseline | 6x lower |
| Break-even detection rate | N/A | 5% | Prevent 1 in 20 failures |