Machine Learning Applications for Vibration-Based Structural Health Monitoring

Introduction

Structural systems, such as transportation, power and energy infrastructure, as well as critical buildings (e.g., schools, hospitals, government buildings) are essential for the functioning of modern societies and their possible failure may have significant consequences. For example, Figure 1 shows some notable examples of bridge collapses around the world. Among the reasons for bridge collapse is their ageing and the increased traffic loads compared to the loads considered at time that the bridge was designed. Environmental reasons, such as flood and scour are also common reasons that cause the collapse of structures which usually were already deteriorated. When the anticipated service life of these structures comes to an end, or when their capacity is reduced, replacing them is quite challenging, or just not possible. Therefore, it is often necessary to monitor the response of an important structure in order to evaluate its health condition and to detect early-stage damage for ensuring operational safety.

Structural Health Monitoring (SHM) involves the observation and the analysis of a system over time using periodically sampled response measurements, or/and digital images in order to monitor changes in the material, or the geometry of the system. The major families of SHM algorithms are vibration-based, and vision-based. Vibration-based SHM can be seen as the traditional approach for monitoring a structural system where sensors (usually accelerometers) are used to measure the structural response under various loading conditions. The collected measurements are then processed in order to extract the dynamic characteristics, or other damage-sensitive features after applying signal processing techniques. Post-processing also aims on extracting meaningful damage indication parameters (damage indices or damage-sensitive features) that will be used to determine structural health. Thanks to the recent technological advances in sensors, drones, high-speed internet, cloud computing, as well as the advances in Machine Learning (ML) techniques, the combination of ML techniques with both SHM families is gaining more and more interest. This Chapter focuses on ML applications for vibration-based SHM offering a detailed and critical examination of the problems and the potential of ML methods.

Figure 1.

Examples of recent bridge collapses: (a) Pittsburg bridge, USA (collapsed 28/01/2022), (b) Mirepoix-sur-Tarn, France (collapse 18/11/2019), (c) Kalambaka, Trikala bridge (collapsed 2016), (d) Plaka bridge, Greece (collapsed 1/02/2015)

Vibration-based SHM can be either model-based (known as physics-based), or data-based (known as data-driven). Model-based SHM usually includes the development of a digital twin of the monitored structure. A digital twin might be a numerical model, an analytical equation, an ML algorithm or a surrogate model, while in model-based SHM, Finite Element (FE) models are commonly used. An example of a structure and its digital twin can be seen in Figure 2. Τhe model properties can be updated through the solution of an inverse problem using the measured data. On the other hand, data-based approaches require a statistical model instead of a digital twin. The development of the statistical model depends on the use of ML or pattern recognition algorithms, and the extracted damage-sensitive features, or on the raw sensor data, since this approach is entirely based upon data. Some researchers have also proposed hybrid SHM approaches which combine the best features of model-based and data-based methods. The major principles of hybrid methods can be found in Barthorpe (2010). It is worth mentioning that in the context of ML applications for SHM, numerical models are used not only for damage detection through model updating but also as a tool for designing an optimal SHM sensor network, and as a proxy for damage induction in order to enable the supervised training of data-based damage detection algorithms (Figueiredo et al., 2019).