Structural health monitoring has been widely employed in several engineering fields as support tool for condition-based maintenance policies aimed at increasing structural safety levels. Among the many applications proposed in the literature, such an approach has been proved to allow accurately characterizing damage in thin-walled structures, which are widespread both in aeronautical and in mechanical applications. In this field, a consolidated solution is represented by tomographic algorithms used to process actively monitored ultrasonic guided waves to generate a probability map of possible damage affecting the structure. More recently, machine learning-based frameworks, specifically neural networks, have been employed as alternative tool to tomographic algorithms to perform damage detection, localization and/or quantification, successfully overcoming some intrinsic limitations of classic methods. However, the black box-like nature of neural networks has built mistrust in such tools, thus creating a gap between their employment in the academic world and in industrial applications. This work aims at reducing such a gap by presenting an explainable machine learning framework for ultrasonic guided wave-based damage diagnosis. Specifically, a convolutional neural network for classification is employed to detect possible damage affecting thin-walled structures. The capabilities of the framework are demonstrated by means of a realistic numerical case study involving crack-like damage affecting a metal plate. Moreover, the behavior of the convolutional neural network is explained through the layer-wise relevance propagation framework. This allows comparing the ultrasonic guided waves features learned by the machine learning algorithm to the intuition of human experts, with the purpose of building trust in the network and, possibly, underlying damage-related physical phenomena hidden to the human eye.
Explainable framework for lamb wave-based damage diagnosis
Lomazzi L.;Giglio M.;Cadini F.
2023-01-01
Abstract
Structural health monitoring has been widely employed in several engineering fields as support tool for condition-based maintenance policies aimed at increasing structural safety levels. Among the many applications proposed in the literature, such an approach has been proved to allow accurately characterizing damage in thin-walled structures, which are widespread both in aeronautical and in mechanical applications. In this field, a consolidated solution is represented by tomographic algorithms used to process actively monitored ultrasonic guided waves to generate a probability map of possible damage affecting the structure. More recently, machine learning-based frameworks, specifically neural networks, have been employed as alternative tool to tomographic algorithms to perform damage detection, localization and/or quantification, successfully overcoming some intrinsic limitations of classic methods. However, the black box-like nature of neural networks has built mistrust in such tools, thus creating a gap between their employment in the academic world and in industrial applications. This work aims at reducing such a gap by presenting an explainable machine learning framework for ultrasonic guided wave-based damage diagnosis. Specifically, a convolutional neural network for classification is employed to detect possible damage affecting thin-walled structures. The capabilities of the framework are demonstrated by means of a realistic numerical case study involving crack-like damage affecting a metal plate. Moreover, the behavior of the convolutional neural network is explained through the layer-wise relevance propagation framework. This allows comparing the ultrasonic guided waves features learned by the machine learning algorithm to the intuition of human experts, with the purpose of building trust in the network and, possibly, underlying damage-related physical phenomena hidden to the human eye.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.