A Cross-validation approach in Neural Networks for fretting-fatigue prediction on Al 7075-T651

This work presents a Feed-Forward Neural Network (FFNN) developed to estimate the fretting fatigue life of aluminium Al 7075-T651 under cylindrical and spherical contact conditions. The study compares two sets of input parameters: one based on easily acquired experimental data (NN1: bulk stress, semi-width of the contact area, and tangential force ratio) and another that incorporates detailed physical characteristics of the phenomenon (NN2: normal stress and strain variations at critical depths, in addition to bulk stress). To optimize training with a limited dataset and mitigate overfitting, the k-fold cross-validation technique was employed. The research evaluated different optimization algorithms, with the Adam optimizer showing superior performance, along with the ReLU activation function. The results demonstrate that neural networks predict fretting fatigue life more accurately than classic methods. While the NN1 model exhibited slightly better performance in a simple data split, the NN2 model suggests a greater capacity for generalization, even for different geometries, due to the inclusion of stresses and strains that can be estimated numerically. K-fold cross-validation proved crucial in improving the reliability of predictions by maximizing the use of available data. In conclusion, neural networks are presented as a promising and accurate tool for predicting fretting fatigue life.