TY - JOUR

T1 - Artificial intelligence assisted fatigue failure prediction

AU - Schneller, Wolfgang

AU - Leitner, Martin

AU - Maier, Bernd

AU - Grün, Florian

AU - Jantschner, Oliver

AU - Leuders, S.

AU - Pfeifer, Tanja

N1 - Publisher Copyright: © 2021 The Authors

PY - 2022/2

Y1 - 2022/2

N2 - This work presents a novel approach for defect based fatigue failure characterization using artificial intelligence (AI). An artificial neural network (ANN) is trained on experimentally determined data that is highly relevant in terms of fatigue. Load stress, hardness and killer defect size are the three main parameters defined as input arguments. Fatigue testing either reveals failure or non-failure, which represent the two possible output variables. Thus, every specimen subjected to this research work generates at least one data set. After total fracture occurs at a certain load level, killer defect size is evaluated by fracture surface analysis. The architecture as well as hyperparameters of the neural network are optimized by K-fold cross validation in order to obtain best prediction accuracy. Eventually, a conservative mean fatigue failure prediction accuracy of 91.6% is achieved. This unprecedented methodology is pioneering to predict fatigue failure without the need for extensive, error-prone, use of complex assessment methodologies and associated comprehensive expensive material testing. Without any expert-knowledge of evaluation procedures, developed AI-approach enables quick and reliable prediction of fatigue failure of machined components based on elementary key figures and shows prospective ways to revolutionize fatigue characterization.

AB - This work presents a novel approach for defect based fatigue failure characterization using artificial intelligence (AI). An artificial neural network (ANN) is trained on experimentally determined data that is highly relevant in terms of fatigue. Load stress, hardness and killer defect size are the three main parameters defined as input arguments. Fatigue testing either reveals failure or non-failure, which represent the two possible output variables. Thus, every specimen subjected to this research work generates at least one data set. After total fracture occurs at a certain load level, killer defect size is evaluated by fracture surface analysis. The architecture as well as hyperparameters of the neural network are optimized by K-fold cross validation in order to obtain best prediction accuracy. Eventually, a conservative mean fatigue failure prediction accuracy of 91.6% is achieved. This unprecedented methodology is pioneering to predict fatigue failure without the need for extensive, error-prone, use of complex assessment methodologies and associated comprehensive expensive material testing. Without any expert-knowledge of evaluation procedures, developed AI-approach enables quick and reliable prediction of fatigue failure of machined components based on elementary key figures and shows prospective ways to revolutionize fatigue characterization.

KW - Artificial intelligence

KW - Fatigue

KW - Keras

KW - Tensorflow

UR - http://www.scopus.com/inward/record.url?scp=85117131488&partnerID=8YFLogxK

U2 - 10.1016/j.ijfatigue.2021.106580

DO - 10.1016/j.ijfatigue.2021.106580

M3 - Article

AN - SCOPUS:85117131488

VL - 155.2022

JO - International Journal of Fatigue

JF - International Journal of Fatigue

SN - 0142-1123

IS - February

M1 - 106580

ER -