A Study on XANNs for Analyzing Failures in Guided Ultrasonic Wave-based Damage Localization

In the field of Structural Health Monitoring (SHM) using Guided Ultrasonic Waves (GUW), machine learning (ML) approaches for data analysis are becoming increasingly popular. Especially, deep learning with artificial neural networks (ANN) has proven effective due to its ability to process a large amount of data. ANN are commonly employed in damage localization tasks, where, with appropriate training, they deliver good results. However, a significant challenge persists in the application of ANN, namely, their "black-box" nature, which hinders the investigation of decision processes within the ANN and makes it difficult to understand why an ANN model fails under certain inputs. In the present work, an already introduced explainable artificial neural network (XANN) approach will be utilized to investigate when and why a trained XANN model fails with given inputs. The Open Guided Wave (GUW) platform's data is employed to train and test a XANN model designed to predict the location of artificial damage. Subsequently, individual inputs are modified (simulating malfunctions in specific sensor paths), and the impact on decision processes within the XANN architecture is examined. The aim of this study is to investigate and enhance the resilience and explainability of ANN approaches in the SHM domain.