A statistical learning approach to response reconstruction for accelerated fatigue testing

Abstract

Accelerated Fatigue Testing (AFT) is the process of performing fatigue testing in order to ascertain the characteristics of a system’s degradation when subjected to repeated events that induce fatigue. In this work the focus is on extreme events in mechanical systems which induce high levels of fatigue. In this work the mechanical system considered are automotive-suspension models. Due to the nature of these events the system’s behaviour is highly nonlinear. To conduct an accurate fatigue test on a system the measured high fatigue incidents must be reproduced in the system in a laboratory environment. This is typically done using servo hydraulic actuators. The amount of data that can be harvested, however, is limited since the testing procedure induces fatigue which compromises the integrity of the AFT test. Performing AFT can be broken down into three steps: sampling the system in order to harvest data, then nonlinear system identification is used to construct a plant model of the system which is then finally used to perform control of the system to trace the target response. This work focuses specifically on the first two items: sampling and nonlinear system identification. In this work an alternative approach to nonlinear system identification is developed based on statistical learning techniques. The approach aims to identify important events in the input and output spaces that characterise the system’s response and construct a mapping between these events by using a high dimensional regression technique developed in this work. This identified regression technique used is Multivariate Principal Component Regression (MPCR). Next, a sampling procedure is developed to find data that are important in order to model the target responses. Therefore, in this work the interface between the sampling and system identification parts of the AFT procedure is investigated. New approaches in terms of sampling and system identification are developed, implemented and compared with current techniques as far as possible. The alternative approach to AFT developed in this work is shown to produce more accurate models of system under study and leads to a more accurate AFT test with higher integrity. Further possible research directions are also addressed.