Fatigue damage in CFRP often develops quietly—until it becomes costly (or unsafe). This paper proposes a hybrid deep-learning approach for fatigue diagnostics in carbon-fiber reinforced polymers using sensor-based structural health monitoring (SHM) signals.
What they did
- 1D-CNN to extract local/spatial degradation patterns from sensor signals
- xLSTM to capture long-range temporal fatigue evolution
- Mutual-information feature selection to remove redundant features
- Bagging ensemble classifier for robust fatigue-state decisions
Key results (NASA-CFRP dataset)
- 99.00% classification accuracy
- Healthy sensitivity: 98.7% (443 correct; 6 misclassified)
- Defective sensitivity: 99.3% (455 correct; 3 misclassified)
- ROC AUC: 0.990129
Why it matters for industry
- Earlier, more reliable fatigue-state detection supports condition-based maintenance
- Fewer missed degradations and fewer false alarms can reduce downtime and lifecycle cost
- A practical pipeline (CNN + sequence model + feature selection + ensemble) can be integrated into existing SHM stacks
How I’d pilot this in 10 business days
- Days 1–3: Collect SHM sensor logs and align labels (healthy/defective) using inspection and maintenance records
- Days 4–7: Train a baseline CNN+xLSTM and validate on a strictly held-out asset/time split
- Days 8–10: Deploy in shadow mode with a simple dashboard, then tune alert thresholds from false alarms and misses