Supervised learning algorithms for damage detection and long term bridge monitoring

Abstract

In the past few years, numerous methods for damage assessment for structural health monitoring were proposed in the literature. Several problems are raised for making these approaches practical for the engineer. The first concern is to determine whether a structure presents an abnormal behavior or not. Statistical inference is concerned with the implementation of algorithms that analyze the distribution of extracted features in an effort to make decisions on damage diagnosis. Learning algorithms have extensively been applied to classification and pattern recognition problems in the past years and deserve to be used for structural health monitoring. In addition, data acquisition campaigns of civil engineering structures can last from several minutes to years. Dealing with large amounts of data is not an easy task and suitable tools are required to correctly extract important features from them: symbolic data analysis (SDA) is such an approach. In this paper, some supervised learning methods (Bayesian decision trees, neural networks and support vector machines) are introduced to discriminate structural features and are developed within the concept of symbolic data analysis in order to compress data without losing its inherent variability. To highlight the different features of these techniques for structural health monitoring, this paper focuses attention on the monitoring of a railway bridge belonging to the high speed track between Paris and Lyon. During the month of June 2003, a strengthening procedure was carried out on this bridge. In so doing, vibration measurements were recorded under three different structural conditions: before, during strengthening. In the following years (2004, 2005 and 2006), new tests were performed to observe how the dynamic behavior of the bridge evolved, especially for the case of frequency changes. The objective was to verify whether the strengthening procedure was still effective or not, in other terms if the new data could be still assigned to the condition .after strengthening.. This paper reports the major results obtained and shows how the supervised learning techniques can be applied to cluster structural behaviors and classify new data. An original assignment approach is also presented: based on dissimilarity measures this approach shows that in fact the structural behavior of the bridge seems totally different of the initial behaviors used for training the supervised learning methods. Supervised learning algorithms