Fatigue life analysis of high-strength bolts based on machine learning method and SHapley Additive exPlanations (SHAP) approach

Abstract

Fatigue failure of high-strength bolts is one of the causes of collapse of steel structures. In this paper, six different machine learning models were used to analyze and predict the fatigue life of high-strength bolts, and the relationship between fatigue life of the bolts and the influencing factors was analyzed by SHAP method. During this process, a data set of fatigue life of high-strength bolts was presented, and 30 percent of the data was randomly selected as the test set. Geometric dimensions and stress states of bolts were as input features, and fatigue life was output label. After training the model, the errors of bolt fatigue life prediction of six machine learning models were compared. Finally, the most adverse factors affecting the fatigue life of the bolts are analyzed. It was found that XGBoost has the best performance on prediction of fatigue life of high-strength bolts, R2 reaches 0.881 and 0.788 in training set and test set, respectively. At the same time, the prediction result of the model is also better than that of the traditional fracture mechanics method. In addition, according to the analysis of SHAP value, the stress amplitude (SA) applied on the bolt has the greatest impact on the fatigue life. Larger SA value will accelerate the expansion of fatigue crack, thus increasing the degree of fatigue damage of the bolt. At the same time, the combination of MAXS (maximum stress applied on high-strength bolts) and SA is the most unfavorable factor to affect the fatigue life of the bolt. Increasing both factors can greatly reduce the fatigue life of the bolt. Finally, the current mainstream steel structure design codes have sufficient safety reserves, and M39 bolt has the best fatigue performance.