A machine learning approach to the prediction of fretting fatigue life

Abstract

The paper analyses some fretting fatigue results from the literature, reported by Nowell and by Szolwinski and Farris. The principal variables of contact size, peak pressure, remote specimen tension, and tangential force ratio are identified and these are used to construct an Artificial Neural Network (ANN), aimed at predicting total fretting fatigue life. The network is trained and validated using 90% of the data, and its success at predicting the results for the remaining 10% of unseen data is examined. The network is found to be very effective at separating the results into low life and ‘run-out’ groups. It is less successful at predicting lives for the low life specimens, but this is largely due to the difficulty of incorporating the runout and finite life tests together in the same dataset. The approach is seen to be potentially useful and identifies contact size as a key variable. However, the results highlight the need for significant numbers of experimental results if the method is to be used effectively in future. Nevertheless, the trained network comprises a useful tool for the prediction of future experimental results with this material.