Transmissibility Functions-Based Structural Damage Assessment with the Use of Explainable Convolutional Neural Networks

Damage assessment in the structural field is an important area of study and research. In this context, the use of vibration data has always created a lot of interest due to the great and cheap availability of sensors, as well as the strong relationship that the gathered signals have with the system damages. Among the different vibration-based features that have been exploited, transmissibility functions (TFs) have shown the potential to solely rely on the system outputs, making their use particularly convenient in practical engineering. However, most of the damage-related signals features are often hidden and to extract them, many pre-processing steps are typically required. With this aim, this paper proposes a new vibration-based structural diagnosis tool that exploits the capabilities of convolutional neural networks (CNNs) to directly extract subtle damage-related features from complex raw TFs spectra. The method is presented for an experimental case study of a steel structure made of beams connected by means of bolted joints. In particular, a CNN is used to accurately localize the damage in the structure, which is introduced by loosening the joint bolts. In addition, to then understand which are the most important input features exploited by the CNN in the damage characterization, an explainable artificial intelligence method, based on the layer-wise relevance propagation algorithm, is also used.