Prediction of construction cable forces of CFST arch bridge based on DNN

Abstract

Identifying an appropriate set of construction cable forces is important for concrete-filled steel-tube (CFST) arch bridges constructed using a cable-based cantilever assembly method. As the span of arch bridge increases, the nonlinear effects become increasingly pronounced, rendering many computational assumptions inadequate and bring the applicability of current methods into question. Additionally, when the actual situation in the construction process deviates from the calculated result, the cable force needs to be adjusted quickly. However, the current method has a long analysis process and does not meet the requirements of rapid construction. Therefore, this paper introduces a Deep Neural Network (DNN)-based method for predicting construction cable forces in large-span CFST arch bridges, aiming to enhance both accuracy and efficiency. A comprehensive nonlinear finite element model was established, incorporating the details of the construction stages and considering various geometric nonlinear effects. Random cable forces were then fed into the model to calculate the corresponding arch rib alignments. Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks are employed to learn from the generated data, forming a nonlinear mapping between arch rib alignments and cable forces, and fully utilizing their capabilities to handle complex and nonlinear problems. The desired construction cable forces corresponding to the objective alignment are subsequently predicted using the DNN model. The proposed method was validated using a large-span CFST arch bridge as a case study, with the results indicating that both the MLP and LSTM predictions meet the construction accuracy requirements and demonstrate remarkable efficiency. In particular, LSTM achieves a higher prediction accuracy owing to its ability to consider the interactions between different cables, making it the recommended choice for complex cable force prediction problems.