Intelligent Aggregation of Single-Sensor Classifiers for Enhanced Structural Health Monitoring Networks
Résumé
Structural health monitoring (SHM) systems for large-scale infrastructures often rely on dense sensor networks, which are prone to faults, generate high-volume data, and require computationally efficient algorithms to ensure low-latency inference for real-time monitoring. To enhance overall network accuracy and robustness, aggregating the predictions of individual sensors provides a way to leverage complementary information across the network while mitigating sensor-level errors. In this study, we investigate intelligent aggregation strategies for singlesensor classifiers in SHM networks. We leverage acceleration time series data from the RT345 bridge dataset, collected from a real instrumented structure, to detect and classify structural damages. Individual sensor classifiers produce probabilistic predictions, which are then combined using different aggregation strategies. Soft averaging serves as a baseline, while stacking ensembles employs linear meta-classifiers (Logistic Regression) for interpretable per-sensor weighting and nonlinear metaclassifiers (Random Forest) to capture complex conditional dependencies across sensors, albeit at the cost of interpretability and stability. Experimental results demonstrate that meta-learning strategies significantly improve classification accuracy and robustness. We further evaluate prediction time, scalability, and model size, highlighting trade-offs between linear and nonlinear aggregation for real-time SHM applications. Finally, we extend the study by exploring alternative acceleration time series representations, showing that system efficiency can be improved without compromising damage detection performance.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |