Neural network survival analysis of concrete bridge decks

Deterioration modeling is a key task in bridge maintenance planning. Advanced deterioration modeling offers ability to focus maintenance action to where and when they are most needed. Accurate predictions of bridge deterioration require extensive historical data about bridge condition development. In the United States, bridge inspection results are compiled into a database known as the National Bridge Inventory (NBI). In the literature, this database has been used in conjunction with statistical approaches for deterioration modeling. In this paper, we use a neural network survival model to model the time bridge decks remain in a given condition rating. Survival analysis is an exciting field of statistics that has been widely used in medical and reliability engineering fields and in bridge engineering. This paper explores a flexible method of modeling bridge deck survival probabilities, aimed at informing the integrity management process. We employ Nnet-survival, a Python library for neural network -based survival analysis originally developed for medical purposes. A neural network approach like the one adopted allows deviation from the traditional proportional hazards assumption. We show how the ability to avoid this assumption is important for bridge condition data. We use the neural network architecture developed to study the temporal variation of relative risk between different bridge characteristics.