Segmental assembly joints with excellent shear bearing capacity are the key to ensure the integrity of precast concrete girder bridges, which directly affects the force transmission state of the girders. However, with the existence of unfavorable factors such as the multi-key shear capacity reduction effect, the traditional prediction model for calculating the concrete joints shear capacity has poor prediction accuracy and large dispersion. To overcome the shortcomings of the existing prediction model, this study established a database consisting of 311 sets of test data and 110 sets of numerical results based on existing research. A total of seven data-driven models for predicting shear capacity of concrete joints were trained and generated based on the database, namely, two linear models (Linear Regression Support Vector Machine Algorithm, Least Squares Linear Regression) and five nonlinear models (Neural Network Bayesian Regularization; Neural Network Quantized Conjugate Gradient Model, Neural Network Levenberg-Marquardt (LM), Decision Tree, and Gaussian regression). The coefficient of determination (R2), root mean square error (RMSE),mean absolute error (MAE), error range (A20-index), and error analysis were adopted to evaluate those model's performance. The evaluation results show that the LM model has excellent prediction accuracy, stability, robustness, and can guide the engineering design. Finally, the established database (421 data sets) and trained LM algorithm model were open sourse in this study to promot the investigate of precast concrete structures.
Read full abstract