The present work investigates the integration of ART with machine learning for damage detection in additively manufactured triply periodic minimal surface (TPMS) lattices. Advanced lightweight components based on lat- tice structures offer superior performance but their complex geometry limits conventional inspection techniques. Acoustic Resonant Testing (ART) provides a non-destructive alternative by capturing changes in dynamic re- sponse that reflect structural integrity of a component. A neural network was developed to detect the presence of damage, identify its location, and estimate its size. Because experimental data were limited, a numerical campaign was conducted to generate a virtual database. The models achieved an elevated accuracy in damage detection. For what concern the damage localization, excellent results were reached, consistently identified the damaged region in post-fatigue specimens, in line with SEM analyses. Size estimation was more challenging, particularly for small defects, reflecting the sensitivity of regression tasks to multiple factors. The results demonstrate the feasibility of combining ART with neural networks for structural assessment of lattice materials, and highlight the potential of this approach as a step toward digital twins for real-time monitoring.
Il presente lavoro analizza l’integrazione dell’ART con tecniche di apprendimento automatico per il rilevamento dei danni in reticoli Triply Periodic Minimal Surface (TPMS), realizzati tramite manifattura additiva. I componenti avanzati a struttura reticolare offrono elevate prestazioni meccaniche, ma la complessità geometrica limita l’efficacia delle tecniche di ispezione convenzionali. L’Acoustic Resonant Testing (ART) rappresenta un’alternativa non distruttiva, in quanto consente di rilevare variazioni nella risposta dinamica associate all’integrità strutturale del componente. È stata sviluppata una rete neurale in grado di individuare la presenza di danni, localizzarli e stimarne le dimensioni. Poiché i dati sperimentali disponibili erano limitati, è stata condotta una campagna numerica per generare un database virtuale. I modelli hanno raggiunto un’accuratezza elvata nel rilevamento dei danni. Anche per quanto riguarda la localizzazione del danno si sono raggiunti ottimi risultati, identificando costantemente la regione compromessa nei provini post-fatica, in accordo con le analisi SEM. La stima delle dimensioni si è rivelata più complessa, soprattutto per difetti di piccola entità, a causa della sensibilità intrinseca dei compiti di regressione a molteplici fattori. I risultati dimostrano la fattibilità di combinare ART e reti neurali per la valutazione strutturale di materiali reticolari e ne evidenziano il potenziale come passo verso lo sviluppo di digital twin per il monitoraggio in tempo reale.
Fatigue damage detection in triply periodic minimal surface (TPMS) structures using acoustic resonance testing (ART) and numerically-informed neural networks
ROSCIOLI, FEDERICO
2024/2025
Abstract
The present work investigates the integration of ART with machine learning for damage detection in additively manufactured triply periodic minimal surface (TPMS) lattices. Advanced lightweight components based on lat- tice structures offer superior performance but their complex geometry limits conventional inspection techniques. Acoustic Resonant Testing (ART) provides a non-destructive alternative by capturing changes in dynamic re- sponse that reflect structural integrity of a component. A neural network was developed to detect the presence of damage, identify its location, and estimate its size. Because experimental data were limited, a numerical campaign was conducted to generate a virtual database. The models achieved an elevated accuracy in damage detection. For what concern the damage localization, excellent results were reached, consistently identified the damaged region in post-fatigue specimens, in line with SEM analyses. Size estimation was more challenging, particularly for small defects, reflecting the sensitivity of regression tasks to multiple factors. The results demonstrate the feasibility of combining ART with neural networks for structural assessment of lattice materials, and highlight the potential of this approach as a step toward digital twins for real-time monitoring.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi_M_Federico_ROSCIOLI.pdf
non accessibile
Descrizione: Tesi completa
Dimensione
21.21 MB
Formato
Adobe PDF
|
21.21 MB | Adobe PDF | Visualizza/Apri |
|
Executive_Summary_Federico_Roscioli_.pdf
non accessibile
Descrizione: Executive summary
Dimensione
1.04 MB
Formato
Adobe PDF
|
1.04 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/243601