Database for training predictive AI-based assessment algorithms in Structural Health Monitoring
Examensarbete för masterexamen
Structural engineering and building technology (MPSEB), MSc
Structural health monitoring (SHM) is a useful technique for ensuring integrity and safety of a structure. Although it is still under development the concept of SHM is to detect possible errors or weaknesses in an early stage and address suitable maintenance or strengthening options. According to Berrocal et al. (2021) a new method in the field of SHM has been established recently by measuring the strain using distributed optical fiber sensors (DOFS). Furthermore, implementing DOFS in reinforced concrete proves to be an accurate and reliable tool for measuring crack widths and deflection. A promising possibility with DOFS is to make automated assessments and predictions of various structural members by combining DOFS with artificial intelligence (AI). The learning capability of AI would be based on a training database comprising the strain data from DOFS and the crack information. Apparently, the economic and environmental cost to produce many different scenarios in a laboratory would be very high. One alternative is to simulate the various scenarios in finite element analysis (FEA) models obtaining artificial strain profiles together with the crack locations. However, this method requires good compatibility between the strain from FE models and DOFS which is not the case. Therefore, postprocessing the artificial strains can hopefully make the two become closely related to each other. If this becomes reality, a future AI training database can be built by collecting strains from FE models with different geometries, material properties, loading conditions and support setups. This thesis aims to examine the problem stated above, i.e., if strains from FE models can, trough manipulation, become comparable to the strains in DOFS. Attempting to answer this question the work was divided in two parts, the calibration procedure, and the strain postprocessing. In the first part, the ambition was to create a digital copy with similar structural behavior as three beams from an experiment made earlier. As a result, a major calibration process took place comparing different random field methods using the JCSS Probabilistic Model Code in concrete. In addition, more detailed comparisons were made between different bond slip characteristics. In the second part, the processing methods were implemented with the intention of synchronising the artificial strains from FE model with the strains from DOFS. Although the FE model had success of demonstrating identical behavior as the tested beams it was hard reproduce the shape from the DOFS. The strains from the DOFS illustrate a rather detailed strain profile, but without the capability of reflecting all the cracks. Under these circumstances it is difficult to reach accordance between the two type of strains. The results indicate that a thinner type of DOFS would perhaps give more satisfying results. This is however left for future research.
Structural health monitoring, distributed optical fiber sensors, finite element analysis, artificial intelligence