Structural Health Monitoring through Computational and Experimental Methods as a Generic Approach to the Damage Detection Problem

Abstract

Detecting damage or changes in structures by monitoring their response has emerged as a field in engineering known as Structural Health Monitoring (SHM). A majority of applications in SHM involve vibration measurements and processing of signals to detect changes in the dynamics related to damage or degradation. The primary points of attention are found, first in the data processing and decision system and second in the data availability for building reliable health prediction models. In the proposed thesis, both problems are tackled in an innovative framework by combining Machine Learning (ML) damage detection and Finite Element (FE) simulations for the necessary training data of ML models. Focus is given on the increasingly popular Artificial Neural Networks (ANN) applied in Deep Learning (DL) form. DL models are able to process and classify signals with little to no pre-processing making them ideal for vibration based SHM tasks. FE models on the other hand can provide arbitrary amounts of simulated vibration responses and generate labelled training data. This is especially useful in cases where access to data of damaged states is not available but may be simulated with an accurate numerical model updated on the experimental intact state. The proposed methodology of FE data generation, DL training on the simulated responses and final validation on corresponding experimental health statuses is tested on three different structures of increasing complexity. The goal of the framework can be summarized in studying the phenomenon of generalizing damage information or patterns in FE simulated signals to the real structures. For the tested cases, detailed studies are presented on how DL model architecture, damage identification problem formulation and FE model error affect the final experimental generalization capability of the proposed SHM framework.