Automatic parameter calibration method for SHM feature generation schemes
Author
Jauregui, Mattin
Term
4. semester
Education
Publication year
2026
Pages
73
Abstract
This thesis investigates how to replace manual, heuristic tuning in a vibration-based SHM feature pipeline with data-driven, statistically grounded, and repeatable procedures. The goal is to automatically calibrate feature-extraction parameters to reliably estimate natural frequencies and damping ratios from operational accelerometer data. Two approaches are developed for automatic spectral band selection: (1) PSD-based peak identification and (2) kurtogram-based analysis of impulsive components (spectral kurtosis). Integrated with the Random Decrement Method (RDM), the approaches are validated on real bridge accelerometer signals from the Loiola viaduct and compared against a reference pipeline. Both approaches converged on using 100 segments of 2 s for RDM averaging, matching the reference settings. Computing efficiency and on-demand energy consumption were assessed to balance accuracy, memory, and execution time; the spectral peak method was faster and more memory-efficient (3.71 s, 0.02 MB) than the kurtogram-based method (8.78 s, 145 MB). System identification (e.g., HAVOK) to derive modal parameters from the RDS is part of the overall process but outside this study’s scope. The findings indicate that automated, data-driven calibration can reduce heuristics and support scalable SHM feature generation.
Denne afhandling undersøger, hvordan man kan erstatte manuel, heuristisk tuning i en vibrationsbaseret SHM-featurepipeline med datadrevne, statistisk begrundede og gentagelige procedurer. Målet er at kalibrere parametre til featureekstraktion, så naturlige frekvenser og dæmpningsforhold kan estimeres pålideligt fra operationelle accelerometersignaler. Der udvikles og afprøves to tilgange til automatisk bestemmelse af relevante spektralbånd: (1) PSD-baseret topidentifikation og (2) kurtogram-baseret analyse af impulsive komponenter (spektral kurtosis). Metoderne indgår sammen med Random Decrement Method (RDM) i en end-to-end pipeline og valideres på virkelige brosignaler fra Loiola-viadukten, hvor de sammenlignes med en eksisterende referencepipeline. Begge tilgange endte med at bruge 100 segmenter af 2 s til RDM-gennemsnit, svarende til referenceopsætningen. Derudover evalueres beregningseffektivitet og on-demand energiforbrug for at afveje nøjagtighed, hukommelse og køretid; den spektrale topmetode var hurtigere og mere hukommelseseffektiv (3,71 s, 0,02 MB) end kurtogrammetoden (8,78 s, 145 MB). Systemidentifikation (fx HAVOK) til udledning af modalparametre fra RDS er del af den samlede proces, men ligger uden for denne studies fokus. Resultaterne viser, at automatiseret, datadrevet kalibrering kan reducere heuristik og understøtte skalerbar SHM-featuregenerering.
[This apstract has been generated with the help of AI directly from the project full text]