Abstract: An improved BP network which use sum of squared relative error (SSRE) as the performance function is applied to predict the fatigue life of three kinds of magnesium alloy under different loading paths. Stain-controlled fatigue experiments are conducted on AZ31B and ZK60 magnesium alloy under four loading paths, which including fully reversed tension-compression, cyclic torsion, 45° in-phase axial-torsion and 90° out-of-phase axial-torsion. In addition, the fatigue data of AZ61A magnesium alloy from literature are also adopted. Two fatigue life prediction methods, namely, a standard BP network which use mean squared error(MSE) as the performance function, and Smith-Watson-Topper(SWT) critical plane fatigue models, are evaluated based on the experimentally obtained fatigue results. Result shows that all of the predicted results except one date by both BP network are within factor-of-three boundaries, there are 16 date, 13 date and 10 date predicted by SWT model outside factor-of-three boundaries even factor-of-five boundaries, respectively. Both BP network are found to be able to correlate the fatigue experiments reasonably well in comparison with SWT model, and SSRE-BP network is better than MSE-BP network.

Key words: BP network, fatigue life, magnesium alloy, sum of squared relative error