A hybrid approach to implement the Digital Twin concept into a damage evolution prediction for composite structures

A hybrid approach to implement the Digital Twin concept into a damage evolution prediction for composite structures

Coenen, Robert (TU Delft Aerospace Engineering)

Zarouchas, D. (mentor)

Delft University of Technology

Aerospace Engineering

2021-11-05

Improvement in the SHM of composite materials requires an enhanced understanding of the damage accumulation processes and helps in the way towards lighter, more optimized, and more sustainable aerospace structures. The Digital Twin concept has the potential to address this problem and may revolutionize the designing, certifying, maintaining, and operating of systems and their components in the long term. This thesis presents a first step to assess its fit in the prediction of damage accumulation within composite materials, specifically the accumulation of transverse matrix cracks within a carbon/epoxy cross-ply specimen under a quasi-static tensile load. To make use of a data-driven technique, a sufficiently large data set is essential. An existing dataset of experiments provides a solid basis but needs augmentation. To augment an existing dataset of tensile tests on cross-plies, a FEM model was constructed that models both transverse matrix cracks and delamination by making use of XFEM-CE. Material variability is implemented per element in the model to overcome the deterministic nature of FEM and generate various crack patterns. The crack patterns and mechanical behavior of the FEM simulations show to be in good agreement with the experimental data. The digital twin that provides real-time predictions is proposed as a LSTM-based neural network. The applied strain to the specimen was predicted with reasonable accuracy. The error of predicting the next crack decreased significantly between predicting the 2nd and 5th crack, after which the mean value of the error and standard deviation remain at similar values. These findings imply that from predicting the 5th crack onward, stochasticity’s role is minimized on the part of predicting the next crack strain. The stochasticity and underlying relationship between the locations of the cracks turned out to be too complex to model with the given approach. Two of the main reasons that are attributed to the latter phenomenon are the modest size of the data set, in spite of further augmentation of the data set after theaddition of the FEM crack patterns, and chosen approach in problem definition.

Digital Twin
composite structure
transverse matrix cracks
LSTM
RNN
damage prediction
XFEM-CE
FEM
deep learning

http://resolver.tudelft.nl/uuid:27e1ac12-5cdb-4154-abcd-3358d244fab4

Student theses

master thesis

© 2021 Robert Coenen