Prediction of welded joint fatigue properties based on a novel hybrid SPDTRS-CS-ANN method

• A prediction model for fatigue properties of welded structures is proposed. • The model is based on SPDTRS-CS-BPNN hybrid algorithm. • A multi-scale fatigue database for EH36 steel is created and the model is applied. • Comparisons with experimental results establish the reliability of the model. • The model can guide the fatigue design of welded structures in industries. This paper proposes a prediction model of welded joint fatigue properties based on single-parameter decision-theoretic rough set (SPDTRS)-cuckoo search (CS)-artificial neural network (ANN) hybrid algorithm. To establish the fatigue performance database of EH36 steel, the preprocessing and data cleaning are carried out by analytic hierarchy process (AHP) and box-plot method to obtain reliable fatigue properties prediction. Therein, the SPDTRS theory is used to analyze the weight of fatigue properties influencing factors, the CS algorithm is used to avoid the over-fitting and local optimization of ANN. During process, the influencing factors are regarded as input and the material related parameters C and m are conducted as output to realize the fatigue properties prediction, which improves the accuracy and stability of the present prediction method. According to the comparisons between the experimental and predicted results, it is found that the predicted S-N curves are within ± 1.1 error band of the experimental results, the average error of fatigue life is less than 10%, and can be within ± 1.2 error band. As a result, the fatigue properties prediction model reasonably shows the fatigue properties of welded structures, and provides a certain reference for fatigue design of welded structures.