Effect of Bridge Data Heterogeneity on Neural Network Survival Predictive Performance

Abstract

Accurate deterioration models are required for data-based bridge management. These models allow converting raw data into actionable insights to guide maintenance decisions. Survival modeling is a technique of modeling time-to-event that has been found helpful for bridge deterioration modeling purposes. Neural network-based survival models have recently shown promise for use in the field. In this work, we investigate the effect of bridge population heterogeneity on the predictive performance of such models. We study this problem in the context of the National Bridge Inventory (NBI), a dataset containing inspection data of US highway bridges. There are many structural systems, materials, deck protection systems, and loading conditions in the NBI bridge population. We hypothesize that this type of heterogeneity will influence the model performance. To test our hypothesis, we split the data into subsets and compare model performance when fitted individually to subsets. In splitting the data, we utilize two separate approaches: statistical clustering and a physics-based approach, where we split the data based on understanding the underlying deterioration mechanisms. By comparing the models fitted to different subsets of data, we can study the effect of data heterogeneity on model performance. The results of this work help further understand the potential limitations the data places on the Neural Network survival model approach. We expect this understanding to improve further development of the modeling approach.