Analysis of the Existing Bridge Stock and Derivation of Design Prior Models

Author: Isabel Stang
Language: English

Abstract

Conceptual design today is primarily based on the experience of the engineering team and the study of available reference projects. This Master’s Thesis investigates the potential of Machine Learning (ML) to assist in the conceptual design of bridges. Different ML models are applied to a dataset of existing bridges with the objective of learning design patterns and creating design prior models.

The study focuses on concrete frame bridges, most of which are single-span rail bridges. First, a cluster analysis is conducted, revealing patterns in the data. The clustering distinguishes rail bridges based on their span, clear height, deck height, and frame type. Road and pedestrian bridges are identified as a separate cluster. Next, various discriminative models (linear model, decision tree, gradient boosted decision tree, neural network) are compared in order to develop a design prior model capable of predicting suitable bridge feature values for a new project situation. The models recognize relationships between parameters that are deemed reasonable for concrete frame bridges. The CatBoost algorithm convinces due to its prediction accuracy, its explainability and its straightforward implementation.

A challenge is posed by the imbalance of the data which, despite the introduction of class weights, leads to inaccurate predictions of underrepresented classes. Furthermore, a generative model, a conditional autoencoder, is examined. This model allows to explore different bridge design alternatives for a new project situation. However, its performance must be further improved to conclude on detected design patterns and apply the model in conceptual design.

Finally, two example project situations of built concrete frame bridges demonstrate the applicability of the three design prior models. While the cluster analysis provides additional insights into the bridge type for the conceptual design, the predictions of the discriminative and generative model serve as a starting point for the design process that is based on bridges constructed in the past. While the discriminative models suggest a single “best-fit” solution, generative models can propose multiple suitable design alternatives.

The results of the study show how data-driven models can provide value in the conceptual design of structures and incentivize the creation of larger, structured databases of existing structures.