A new approach to damage detection in bridges using machine learning

Abstract

At the same time that civil engineering structures are increasing in number, size and longevity, there is a conforming increasing preoccupation regarding the monitoring and maintenance of such structures. In this sense the demand for new reliable Structural Health Monitoring systems and damage detection techniques is high. A model-free damage detection approach based on Machine Learning is presented in this paper. The method performs on the collected feature measurements on a railway bridge, which for this study were gathered in a numerical experiment using a three dimensional finite element model. The first step of the approach consists in collecting the dynamic response of the structure, simulated during the passage of a train over the bridge, in both the healthy and damage states of the structure. The next step consists in the design and unsupervised training of Artificial Neural Networks that use as input accelerations and axle loads and compute a novelty index, called prediction error, based on a novelty detection approach. The distribution of the obtained prediction errors is statistically evaluated by means of a Gaussian Process and, after this process, damage indexes can be defined. Finally, the efficiency of the method is assessed in terms of Type I (false positive) and Type II (false negative) errors using Receiver Operating Characteristic curves. The promising results obtained in the case study demonstrate the capability of the presented method.