Deterioration modeling is a critical task in bridge maintenance planning. Advanced deterioration modeling allows maintenance actions to focus on where and when they are most needed. In this paper, we use a neural network survival model applied to National Bridge Inventory (NBI). Survival modeling is a technique of modeling the time-to-event and has been traditionally associated with fields such as medical sciences and reliability engineering. In this work, we focus on understanding the effect of bridge population heterogeneity on model performance. Multiple structural systems, materials, and deck protection methods exist within the bridge population. In addition, bridges are subjected to various environmental and loading conditions. We expect that heterogeneity of the population will provide difficulties in selecting a suitable model. To understand this problem, we study the effect of heterogeneity in the bridge population on model prediction performance. To do this, we first split the dataset into subsets. We approach the problem of dividing the data from two unique angles: statistical clustering methods and a physics-based approach where we split the dataset into subsets based on the underlying deterioration mechanisms. After splitting the data into subsets, we fit separate survival models on each of these subsets and compare the prediction performance with the other subset models and a model fitted on the entire dataset. This comparison allows us to understand if the type of survival models we have utilized is more suitable for some deterioration mechanisms and structural types than others.
|