The field of prognostics on composites is relatively young, and research is focused on constant amplitude fatigue (CAF) loading, whereas variable amplitude fatigue (VAF) loading is more common in actual use-cases. Therefore in this research, the feasibility of different in-situ, data-driven probabilistic models is studied for prognostics on carbon fibre reinforced polymer (CFRP) specimens under VAF. Fatigue data with recorded acoustic emissions (AEs) is available from an earlier performed experimental campaign. Three models were selected: a statistical model to be used as a baseline comparison, a Gaussian process (GP) regression using cumulative AE energy, and a recurrent neural network (RNN) using all available AE features and load data.
The models were compared for performance of remaining useful life (RUL) predictions on specimen under VAF loading for three different cases; when trained on CAF, VAF, and the combination of these two. Seven performance metrics were used to quantify their performance, as well as a qualitative comparison.
The statistical method's performance varied per test specimen and can, therefore, only be used in practical applications when used very conservatively. It did not perform better in any of the three training cases as compared to the others.
The GP regression was deemed not feasible due to high variability in its RUL predictions, high uncertainty due to the setting of a failure threshold as probability distribution based on other specimens, and high computational costs. The performance differed per test specimen as well.
Finally, the performance of the RNN increased when trained on VAF data as compared to training on solely CAF data. It increased further when including both CAF and VAF in the training data. It is not yet feasible to be used in practice, due to variability in its predictions, and the inability to handle outliers. The latter is an issue for the other two models as well.
The RNN outperformed the other two models when trained on VAF, and CAF and VAF data. Due to the definition of the failure threshold in the GP, the GP did not perform better than the statistical model. In the case of CAF training data, there was not a clear distinction between the performance of the statistical model and the RNN, except for one performance metric related to the precision of predictions.
During this research, AE data from glass fibre reinforced polymer specimens tested on tension-tension fatigue under different load levels became available. In a case study, the feasibility of a RNN, trained on AE data, was analysed for prognostics on this data-set. Due to large differences in the life-times between the specimens in this data-set, this was not feasible. Extending this RNN with a feedforward neural network which uses load data as well as input, provided worse predictions.
The main conclusion drawn in this thesis is that with the current implementation of the used models, in-situ, data-driven prognostics on composite specimens under VAF is not yet feasible.