Evaluation of time series forecasting approaches for the reliable crack length prediction of riveted aluminium plates given insufficient data

In all fields, the significance of a reliable and accurate predictive model is almost unquantifiable. With deep domain knowledge, models derived from first principles typically outperforms other models in terms of reliability and accuracy. When it may become a cumbersome or an unachievable task to build or validate such models of complex (non-linear) systems, machine learning techniques are employed to build predictive models. However, the accuracy of such techniques is not only dependent on the hyper-parameters of the chosen algorithm, but also on the amount and quality of data. This paper investigates the application of classical time series forecasting approaches for the reliable prognostics of technical systems, where black box machine learning techniques might not successfully be employed given insufficient amount of data and where first principles models are infeasible due to lack of domain specific data. Forecasting by analogy, forecasting by analytical function fitting, an exponential smoothing forecasting method and the long short-term memory (LSTM) are evaluated and compared against the ground truth data. As a case study, the methods are applied to predict future crack lengths of riveted aluminium plates under cyclic loading. The performance of the predictive models is evaluated based on error metrics leading to a proposal of when to apply which forecasting approach.