Lamb waves excited by a network of piezoelectric transducers have been used in the recent past for Structural Health Monitoring (SHM) as a means to detect, localize and quantify damages within a structure, given their high sensitivity to the properties of the media they travel through. Moreover, piezoelectric transducers are inexpensive, compact and light, making them suitable for installation in aircrafts, where volume and weight constraints are significant. However, the signals involved are complex and require heavy pre-processing in order to extract useful results, due to the amount of possible signal variation sources that are hard to account for and the limited portion of the signals that can be used for diagnostics. Convolutional Neural Networks (CNNs) are capable of working with these types of signals without the loss of information from pre-processing operations and develop models that successfully localize the damaged portion of the structure. Furthermore, they can be used as autoencoders, eliminating the need for labeled data and working as an unsupervised machine learning algorithm. In this work, a methodology based on Lamb waves and CNN autoencoders for damage localization is proposed. The method is validated with reference to a numerical study case based on an aluminum plate and an empirical study case based on a full scale composite wing panel, proving its generalisation capability and accuracy.
Le Lamb waves eccitate da una rete di trasduttori piezoelettrici sono utilizzate nell’ambito della Structural Health Monitoring (SHM) per rilevare, localizzare e quantificare i danni all'interno di una struttura sfruttando la loro elevata sensibilità alle proprietà dei mezzi che attraversano. Inoltre, essendo i trasduttori piezoelettrici economici, compatti e leggeri risultano ideali per applicazioni aeronautiche dove i vincoli di volume e di peso sono particolarmente restrittivi. Tuttavia, l’elevata quantità di possibili fonti di disturbo dei segnali e la limitata porzione degli stessi utilizzabile per la diagnostica rendono complessa la correlazione fra i disturbi e le fonti richiedendo un processo consistente di preelaborazione per poter ricavare risultati utili. Con le reti neurali convoluzionali (dall'inglese Convolutional Neural Network, CNN) risulta possibile sviluppare modelli per la localizzazione della parte danneggiata della struttura senza la necessità del processo di preelaborazione. Inoltre, possono essere utilizzate come autocodificatori, eliminando la necessità di dati etichettati e lavorando come un algoritmo di apprendimento automatico non supervisionato. In questo lavoro viene proposta una metodologia basata sull’interazione fra le Lamb waves e le CNN autoencoders per la localizzazione dei danni interni alle strutture. Il metodo è validato con riferimento a un caso di studio numerico basato su una piastra di alluminio e ad un caso di studio empirico basato su di un pannello alare, in scala reale, realizzato in materiale composito, dimostrando capacità di generalizzazione e accuratezza.
Structural health monitoring using guided waves and unsupervised machine learning
JUNGES, RAFAEL
2020/2021
Abstract
Lamb waves excited by a network of piezoelectric transducers have been used in the recent past for Structural Health Monitoring (SHM) as a means to detect, localize and quantify damages within a structure, given their high sensitivity to the properties of the media they travel through. Moreover, piezoelectric transducers are inexpensive, compact and light, making them suitable for installation in aircrafts, where volume and weight constraints are significant. However, the signals involved are complex and require heavy pre-processing in order to extract useful results, due to the amount of possible signal variation sources that are hard to account for and the limited portion of the signals that can be used for diagnostics. Convolutional Neural Networks (CNNs) are capable of working with these types of signals without the loss of information from pre-processing operations and develop models that successfully localize the damaged portion of the structure. Furthermore, they can be used as autoencoders, eliminating the need for labeled data and working as an unsupervised machine learning algorithm. In this work, a methodology based on Lamb waves and CNN autoencoders for damage localization is proposed. The method is validated with reference to a numerical study case based on an aluminum plate and an empirical study case based on a full scale composite wing panel, proving its generalisation capability and accuracy.File | Dimensione | Formato | |
---|---|---|---|
Tesi_Rafael_Junges__Structural_health_monitoring_using_guided_waves_and_unsupervised_machine_learning__Politecnico_di_Milano.pdf
accessibile in internet per tutti
Dimensione
7.02 MB
Formato
Adobe PDF
|
7.02 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/10589/186655