Performance optimization of BP-DNN prediction model of suction caisson uplift bearing capacity employing modified Co-teaching method

Abstract

In recent decades suction caisson has received increasing attention in offshore deepwater geotechnical foundation solutions. Accurately predicting the uplift bearing capacity of suction caissons is of significant importance in practical engineering. The objective of this research is to propose a more optimized prediction model using backpropagation deep neural network (BP-DNN) by improving the data quality using a modified Co-teaching denoising and fusion method. The database was built which contains a large number of results by experimental and numerical research from literature. Due to the variability of numerical results, the BP-DNN prediction model based on numerical data has greater error than that based on experimental data. Therefore, the Co-teaching denoising method was modified and then adopted to filter and obtain relatively high-quality numerical data. Then the optimal fusion model was developed using the data sampling plan with 2/3 experimental data and 1/3 experimental data + all clean numerical data. The overall performance of the fusion model was proved to be satisfactory.