Genetic algorithm-based fatigue data editing technique

Abstract

Durability testing is an essential process node in automotive component design analysis. The test associated with loading history can be accelerated if the fatigue data editing approach is considered for simplifying the given history. Even though they have been proven, unfortunately, the existing editing techniques involve complex mechanisms (e.g. abrupt detection, Fourier transformation and wavelet analysis), which are complicated in nature and which demand high computational costs. Therefore, this paper is a proposal of a simple technique that makes use of the rule-based fatigue segment classifier when deciding which parts of the history need to be removed. Rules representing the new labelling practice have been generated based on the classification data mining framework. In the context of this study, a rule set represents a group of undiscovered relationship between time domain statistical parameters and damage level. A dataset consisting of an equal length of fatigue segments trained using a multi-objective approach called the Elitist Non??dominated Sorting in Genetic Algorithm (NSGA-II) for seeking several optimal sets of rules (i.e., classifiers) by maximizing predictive accuracy and comprehensibility. The number of attributes underlying the rule set is referred to for final classifier selection where the fitter solution serves as the proposed editing technique. Comparison results on strain-stress cycle properties for the edited history and the full-length version shows that the proposed technique is suitable for fatigue data editing. Moreover, it has an additional benefit that no prior requirement on the frequency or time frequency analysis is needed, providing the damage level of fatigue segments rapidly and the discovering of linguistic knowledge as a novelty.