From Binary to Multi-Class: Neural Networks for Structural Damage Classification in Bridge Monitoring Under Static and Dynamic Loading

• Published in: Dynamics 2024-10, Vol.4 (4), p.786-803
• Main Authors: Kardoulias, Andreas, Arailopoulos, Alexandros, Seventekidis, Panagiotis
• Format: Article
• Language: English
• Subjects: artificial neural networks (ANNs); convolutional neural networks (CNNs); deep neural networks (DNNs); FEM; machine learning (ML); structural health monitoring (SHM)
• ISSN: 2673-8716; 2673-8716
• DOI: 10.3390/dynamics4040040

Structural Health Monitoring (SHM) plays a vital role in ensuring the health status of a wide range of structures, such as bridges, buildings, and large infrastructure in general. The advantages of this process can be further enhanced by incorporating more numerical and statistical approaches into traditional methods, such as finite element analysis and Machine Learning. In this study, a truss bridge structure is examined, and neural networks are trained with data derived from finite element analyses under static loads and dynamic excitations. The contributions of this work are based on comparing neural networks trained with static and dynamic analyses, as well as deriving important insights into the key parameters that impact their performance in SHM. Initially, a binary classification problem is addressed, where numerically trained classifiers are tasked with identifying whether the structure is in a healthy state or not. This category is further divided into two subcategories, depending on the extent of the damage present in the structure. Subsequently, a multi-class classification problem is defined, where three different damage classes of the same extent are considered, and the trained network is required to distinguish between them. Although the training of all neural networks was highly satisfactory, the prediction results varied, with success rates ranging from 55% to 90%. Finally, conclusions are drawn from the results of the study regarding the model error influence, the impact of the damage size, and the types of neural networks and training data used.