Among the maintenance policies adopted to guarantee the safety of structures throughout their service life, condition-based maintenance policies driven by structural health monitoring approaches have progressively gained importance over the last years. Within this field, among the several methods proposed in the literature to diagnose damage affecting thin-walled structures, satisfactory performances have been achieved by adopting tomographic algorithms to process ultrasonic guided waves, even though with some limitations. Recently, such limitations have been overcome by adopting machine learning-based algorithms, even though their implementation in real life applications is still limited because of the mistrust in neural networks determined by their black box-like nature. To date, however, several explainability algorithms have been proposed to interpret the behaviour of neural networks, in particular in the medical and in the military fields, where trust in the tools adopted must be guaranteed. Thus, exploiting the potentialities of explainability frameworks, in this work the layer-wise relevance propagation algorithm is employed to explain the predictions of convolutional neural networks for classification and for regression that characterise damage by processing ultrasonic guided waves excited and sensed by means of a 2-D network of piezoelectric devices. First, the explainability algorithm is applied to give a score to each sample of the acquired ultrasonic guided waves, then such scores are collected by means of a properly developed aggregation strategy to rank the most informative couples of piezoelectric devices. The capabilities of the explainable damage diagnosis framework are demonstrated by means of a numerical, yet realistic, case study involving a metal plate affected by crack-like damage. In particular, the focus is set on the explanation of the behaviour of the neural networks involved, with the aim of building trust in such algorithms and, possibly, revealing damage-related hidden features of ultrasonic guided waves.
On the explainability of convolutional neural networks processing ultrasonic guided waves for damage diagnosis
Lomazzi L.;Parziale M.;Giglio M.;Cadini F.
2023-01-01
Abstract
Among the maintenance policies adopted to guarantee the safety of structures throughout their service life, condition-based maintenance policies driven by structural health monitoring approaches have progressively gained importance over the last years. Within this field, among the several methods proposed in the literature to diagnose damage affecting thin-walled structures, satisfactory performances have been achieved by adopting tomographic algorithms to process ultrasonic guided waves, even though with some limitations. Recently, such limitations have been overcome by adopting machine learning-based algorithms, even though their implementation in real life applications is still limited because of the mistrust in neural networks determined by their black box-like nature. To date, however, several explainability algorithms have been proposed to interpret the behaviour of neural networks, in particular in the medical and in the military fields, where trust in the tools adopted must be guaranteed. Thus, exploiting the potentialities of explainability frameworks, in this work the layer-wise relevance propagation algorithm is employed to explain the predictions of convolutional neural networks for classification and for regression that characterise damage by processing ultrasonic guided waves excited and sensed by means of a 2-D network of piezoelectric devices. First, the explainability algorithm is applied to give a score to each sample of the acquired ultrasonic guided waves, then such scores are collected by means of a properly developed aggregation strategy to rank the most informative couples of piezoelectric devices. The capabilities of the explainable damage diagnosis framework are demonstrated by means of a numerical, yet realistic, case study involving a metal plate affected by crack-like damage. In particular, the focus is set on the explanation of the behaviour of the neural networks involved, with the aim of building trust in such algorithms and, possibly, revealing damage-related hidden features of ultrasonic guided waves.File | Dimensione | Formato | |
---|---|---|---|
1-s2.0-S0888327022007282-main.pdf
Accesso riservato
Dimensione
5.7 MB
Formato
Adobe PDF
|
5.7 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.