A large number of different techniques are used for structural health monitoring purposes. In this thesis the physical phenomenon of Lamb waves is exploited, this being elastic waves that propagate in solid materials of small thickness (i.e. plates). Specifically, a 500x500x5 mm steel plate was simulated using a Finite Element software, within which Lamb waves were excited by means of a simulated piezoelectric sensor placed in the center of the plate. A second simulated piezoelectric sensor was positioned at various locations of the plate and was used as receiver. The signal acquired by the second sensor comprises both the direct arrival of the waves transmitted by the actuator and the reflections of the latter from the edges of the plate and from any damage introduced inside the plate itself. Reading and interpretation of the acquired signals was performed via three different sets of Machine Learning (ML) algorithms: multilayer perceptron (MLP), Long Short-Term Memory (LSTM) and WaveNet (an algorithm based on deep convolutional neural networks recently developed for acoustic applications). The use of a numerical model was essential to generate the high number of examples necessary to train the tested algorithms. The dataset on which the analyses were performed contains signals acquired by the simulated receiver placed in three different locations of the plate, in both undamaged (baseline case) and damaged conditions (obtained by adding a through-thickness defect at 100 different positions), both for temperatures varying from 20°C to 69°C at steps of 1°C (50 temperatures). The instrumentation noise effect have been included into the signals with various levels of intensity, to make them more realistic. The temperature effect was partially compensated using the Baseline Signal Stretch (BSS) technique, the only signal pre-processing tool required for the methodology discussed in this thesis. The main task of this work was to determine the hyperparameters for each type of ML algorithm that would give the best defect detection performance. WaveNet was proven to be the best algorithm for the task when considering computation speed, ease in hyperparameters tuning and accuracy of defect detection, which reached 100% in a number of cases. WaveNet was then tested in two other contexts. The first concerned the robustness of the algorithm in evaluating signals acquired at temperatures outside the range used for its training: the measured performance was largely satisfactory, with 100% detection accuracy reported for signals acquired up to 30°C beyond the training range. The second aimed to validate its performance when applied to a real dataset. The experimental set was acquired under a similar configuration to that used in the numerical simulations, although the plate was made of composite material and the excitation frequency was significantly lower. However, in some cases the algorithm was able to completely discern the signals with defect from those without defects, hence suggesting a strong adaptation capability that can be exploited in different applications.
Un gran numero di tecniche differenti viene utilizzato nell’ambito del monitoraggio strutturale. In questa tesi viene sfruttato il fenomeno fisico delle onde di Lamb, onde elastiche che si propagano in materiali solidi di piccolo spessore (i.e. piastre). In particolare, una piastra in acciaio da 500x500x5 mm è stata simulata utilizzando un software agli elementi finiti, all'interno del quale le onde di Lamb sono state eccitate per mezzo di un sensore piezoelettrico simulato posto al centro della piastra. Un secondo sensore piezoelettrico simulato è stato collocato in varie posizioni della piastra e utilizzato come ricevitore. Il segnale acquisito dal secondo sensore comprende l'arrivo diretto delle onde trasmesse dall'attuatore e le loro riflessioni date dai bordi della piastra e da qualsiasi danno introdotto all'interno della piastra stessa. La lettura e l'interpretazione dei segnali acquisiti sono state affidate a tre diversi set di algoritmi di Machine Learning (ML): multilayer perceptron (MLP), Long Short-Term Memory (LSTM) e WaveNet (un algoritmo basato su reti neurali convoluzionali recentemente sviluppato per applicazioni acustiche). L'uso di un modello numerico è stato essenziale per generare l'alto numero di esempi necessari ad addestrare gli algoritmi testati. Il dataset su cui sono state eseguite le analisi contiene segnali acquisiti dal ricevitore simulato posto in tre diverse posizioni sulla piastra, sia in condizioni non danneggiate (caso base) che in condizioni danneggiate (ottenute con l’aggiunta di un difetto passante collocato in 100 posizioni differenti), entrambe per temperature che variano tra i 20°C ed i 69°C con step di 1°C (50 temperature). L'effetto del rumore della strumentazione è stato incluso nei segnali con vari livelli di intensità, per renderli più realistici. L'effetto della temperatura è stato parzialmente compensato utilizzando la tecnica Baseline Signal Stretch (BSS), unico strumento di pre-processing dei segnali richiesto per la metodologia discussa in questa tesi. Il compito principale di questo lavoro è stato determinare gli iperparametri per ogni tipo di algoritmo di ML che avrebbero dato le migliori prestazioni nel rilevamento dei difetti. È stato dimostrato come WaveNet sia l'algoritmo migliore per tale scopo considerando velocità di calcolo, facilità nella messa a punto degli iperparametri e accuratezza nel rilevamento dei difetti, la quale ha raggiunto il 100% in diversi casi. WaveNet è stato quindi testato in altri due contesti. Il primo riguardava la robustezza dell'algoritmo nella valutazione di segnali acquisiti a temperature al di fuori dell'intervallo utilizzato per il suo allenamento: le prestazioni misurate sono state ampiamente soddisfacenti, con un'accuratezza nel rilevamento del 100% ottenuta per segnali acquisiti fino a 30°C oltre l'intervallo di allenamento. Il secondo mirava a convalidare le sue prestazioni quando applicato ad un set di dati reale. Il set sperimentale è stato acquisito con una configurazione simile a quella utilizzata nelle simulazioni numeriche, sebbene la piastra fosse realizzata in materiale composito e la frequenza di eccitazione fosse significativamente inferiore. Tuttavia, in alcuni casi l'algoritmo è stato in grado di discernere completamente i segnali con difetto da quelli senza difetti, suggerendo quindi una forte capacità di adattamento che può essere sfruttata in diverse applicazioni.
Comparison of MLP, RNN and CNN machine learning algorithms for damage detection exploiting ultrasonic Lamb waves
Urbani, Matteo
2019/2020
Abstract
A large number of different techniques are used for structural health monitoring purposes. In this thesis the physical phenomenon of Lamb waves is exploited, this being elastic waves that propagate in solid materials of small thickness (i.e. plates). Specifically, a 500x500x5 mm steel plate was simulated using a Finite Element software, within which Lamb waves were excited by means of a simulated piezoelectric sensor placed in the center of the plate. A second simulated piezoelectric sensor was positioned at various locations of the plate and was used as receiver. The signal acquired by the second sensor comprises both the direct arrival of the waves transmitted by the actuator and the reflections of the latter from the edges of the plate and from any damage introduced inside the plate itself. Reading and interpretation of the acquired signals was performed via three different sets of Machine Learning (ML) algorithms: multilayer perceptron (MLP), Long Short-Term Memory (LSTM) and WaveNet (an algorithm based on deep convolutional neural networks recently developed for acoustic applications). The use of a numerical model was essential to generate the high number of examples necessary to train the tested algorithms. The dataset on which the analyses were performed contains signals acquired by the simulated receiver placed in three different locations of the plate, in both undamaged (baseline case) and damaged conditions (obtained by adding a through-thickness defect at 100 different positions), both for temperatures varying from 20°C to 69°C at steps of 1°C (50 temperatures). The instrumentation noise effect have been included into the signals with various levels of intensity, to make them more realistic. The temperature effect was partially compensated using the Baseline Signal Stretch (BSS) technique, the only signal pre-processing tool required for the methodology discussed in this thesis. The main task of this work was to determine the hyperparameters for each type of ML algorithm that would give the best defect detection performance. WaveNet was proven to be the best algorithm for the task when considering computation speed, ease in hyperparameters tuning and accuracy of defect detection, which reached 100% in a number of cases. WaveNet was then tested in two other contexts. The first concerned the robustness of the algorithm in evaluating signals acquired at temperatures outside the range used for its training: the measured performance was largely satisfactory, with 100% detection accuracy reported for signals acquired up to 30°C beyond the training range. The second aimed to validate its performance when applied to a real dataset. The experimental set was acquired under a similar configuration to that used in the numerical simulations, although the plate was made of composite material and the excitation frequency was significantly lower. However, in some cases the algorithm was able to completely discern the signals with defect from those without defects, hence suggesting a strong adaptation capability that can be exploited in different applications.| File | Dimensione | Formato | |
|---|---|---|---|
|
Master Thesis Matteo Urbani - 899699.pdf
accessibile in internet per tutti
Descrizione: Tesi Magistrale
Dimensione
9.21 MB
Formato
Adobe PDF
|
9.21 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/10589/167301