Structural Health Monitoring of Bridges
Data-based damage detection method using Machine Learning
Time: Fri 2020-10-16 14.00
Subject area: Civil and Architectural Engineering, Structural Engineering and Bridges
Doctoral student: Ana C. Neves , Bro- och stålbyggnad
Opponent: Associate Professor Elsa Caetano, University of Porto, Department of Civil Engineering
Supervisor: Professor Raid Karoumi, Bro- och stålbyggnad, Järnvägsgruppen, JVG, Byggkonstruktion
Civil engineering structures built according to modern codes are designed for a service life of normally more than 100 years. At the same time, there is a growing pressure to keep existing aged structures in service despite the fact that they have reached the original designed lifetime, with bridges being a good example of this. Naturally, in order to meet these goals, one must ensure the safety of these vulnerable structures by improving their reliability, as well as safeguarding their users. While doing so, infrastructure owners and managers have the desire to maintain the costs as low as possible. One of the key objectives is to identify damage in the structure at its earliest possible stage so that, if deemed appropriate, intervention can take place promptly. For the exposed reasons, there is an upsurge in the demand for clever strategies that can support maintenance and decision-making concerning structures. These are based on cost-effective and reliable inspection and monitoring solutions defined through a process commonly referred to as Structural Health Monitoring. This thesis consists of an extended summary and four appended papers. The author proposes an approach to Structural Health Monitoring of bridges with focus on damage detection, by means of applying data-based methods and techniques for statistical analysis. The research work starts with the development of the method in Paper I, validated with data generated from a simple numerical model of a railway bridge. In this and subsequent papers with respective case studies, data is collected during train passages, in healthy and damaged states of the structure. Part of the data collected under the healthy state is used for training Artificial Neural Networks, as the primary algorithm of the proposed method. Afterwards, new data collected under healthy or damaged states can be directly compared with the correspondent predictions by the networks. The research work then follows logically into assessing the cost of implementing SHM in Paper II. More specifically, the expected cost associated with possible alternatives regarding maintenance decision-making is assessed in light of a case study and its variants. Then, in Paper III, the influence of the frequency content of the collected data on the performance of the networks is investigated. It mainly deals with revealing that the information contained in the higher frequencies of the measured response is not only non-negligible for the purpose of damage detection but actually more useful than the information extracted from the lower frequencies. This conclusion is verified in light of both numerical and experimental field data. Finally, in Paper IV, the proper use of Artificial Neural Networks for the assessment of structural condition is given more attention. This is put into effect by investigating which adjustments in the hyperparameters, features and data structure end up enhancing the performance of the algorithm. The results presented in this thesis demonstrate the potential benefits of integrating the proposed approach into current Structural Health Monitoring practices. This research work intends to contribute with recommendations for proactive maintenance by which data is continuously collected and analyzed in almost real-time, providing the grounds for well-informed decision-making.