Feasibility Study of the Development of a Digital Twin for the Structural Health Monitoring of Marine Structures Using Big Data

Abstract

Structural health monitoring of maritime and offshore structures can contribute to the reduction of the amount of uncertainty in assessment of the operational loads and the structural integrity. Such a survey can also reduce the large safety margins during the design and optimize the conservative inspection and maintenance schedules. One of the emerging variants of structural health monitoring are the so-called digital twin systems that not only allow the monitoring of the structures, but can also simulate the past or the future state of the structural integrity. In the context of this thesis, the feasibility of a digital twin system for the structural health assessment of marine structures based on machine learning algorithms and utilization of big data is examined. The main research question is if the digital twin technology can be enhanced with machine learning to accurately monitor the structural health of marine structures in terms of loading and fatigue damage accumulation. In order to answer this question four modules have been developed and tested. The first module uses artificial neural networks trained on operational data to predict the fatigue damage accumulation and the frequency that corresponds to the maximum stress power density. The error of the developed networks turned out to be negligible in the investigated cases, with standard deviation of less than 4% in the predictions. The peak frequency of the stress power density is predicted using a random forest regression algorithm. The optimized version of the algorithm has led to an accuracy of 93% with standard deviation of less than 1% in the predictions.
The second module is used to recalibrate the design response amplitude operators of the structure using the predictions of the first module or the operational data, if available. The design response amplitude operators are scaled and shifted in order to minimize the deviation between the spectral fatigue calculations and the predicted (or measured) data.The third module is a static/quasi-static load-reconstruction module. Using on-board strain measurements it is able to calibrate the loading of the structure based on a reformulation of the conventional finite element problem for static and quasi-static loading. A sensitivity analysis of the algorithm effectiveness have been performed using multiple load cases in which the reconstruction error turned out negligible. The fourth and final module relates to structural reliability analysis. This module is based on the estimation of the Hansofer-Lind reliability index as a minimization problem. The employed optimization engine is a variation of the particle swarm optimization algorithm using chaotic system behaviour to iteratively calculate the user-defined parameters. The limit state equation is formulated in a way that it takes the uncertainties related to the fatigue damage accumulation prediction model and Miner's Rule into consideration. Uncertainties related to the material and the fabrication process are not taken into account. The developed modules have been tested on structural details of a vessel monitored in the scope of Monitas Joint Industry Project (JIP). About a year's worth of data has been used to train the machine learning algorithms.