Improving the sensitivity of damage detection using structural health monitoring and pattern recognition

Abstract This present thesis focuses on Structural Health Monitoring (SHM), defined as a process to develop strategies in order to identify an abnormal behavior such as a damage, in structural system. SHM can be considered under two approaches, namely the Inverse and Forward. The first one is associated with numerical models, while the second one with artificial intelligence and pattern recognition. The advantage of computational simplicity motivated the use in this thesis of forward SHM approach. The forward SHM approach was applied to the chosen case study, the Overpass Z24, located in Switzerland, which was subject to a controlled demolition. In this dissertation, methods were used to define relations between the structural actions, such as temperatures, and the responses, acquired from the sensors, such as modal frequencies. By exploiting these models, it was possible to analyze the existence of unexpected behavior or structural changes such as damages. The developed models consist of regression models, namely multivariate linear regression, multivariate bilinear regression and the Neural Networks. It was concluded that the best method that can highlight damage is the Neural Network. After conducting this type of modelling, a further analysis on the Neural Network training was developed. A period of data defined as training period was used to create a series of Neural Networks models and another one called testing period to verify the most reliable one and find the best solutions. After fixing The Best Neural Network Model for the case study, several damage indicators, were defined, based on probabilities, clustering and auto correlation of the Neural Networks Errors. The first relies on the definition of statistical boundary conditions, considering the hypothesis of normal distribution, while the second indicator is independent from the creation of boundary conditions, and relies only on the comparison of data distributions. The distributions analyzed consist of the residuals errors obtained for Neural Network estimations. The third and fourth damage indicators consist of the distance between clusters, using the clustering method k-means, and the average auto-correlation function. The following step was to apply those indicators to a scheme of moving data windows defined over time in order to see how they develop when the damage occurs. This method was considered in four different approaches. All indicators were applied using each approach and the results were compared. It can be concluded that for two approaches the results are better, while the others have not shown significant results. Finally, a combination of the indicators studied and applied was used to define new improved damage indicators with superior sensitivity and reliability.