Bayesian neural networks for large-scale infrastructure deterioration models

Abstract

State-space models (SSM) have been shown to be effective at modelling structural deterioration of transportation infrastructure based on visual inspections. The SSM approach was recently coupled with kernel regression (KR) to include structural attributes like age and location in the deterioration analysis to share information between similar structures. However, the existing SSM-KR method suffers from two major drawbacks: 1) it can only use a limited number of structural attributes and 2) it requires significant computational time and resources. This paper proposes a new method, titled SSM-TAGI, that uses a Bayesian neural network instead of KR for extracting information from structural attributes. The new SSM-TAGI approach is compared against SSM-KR using visual inspection data and structural attributes from a network of bridges in Canada. The new SSM-TAGI approach is shown to reduce the computational time by two orders of magnitude while maintaining comparable performance as measured by the test-set log-likelihood. SSM-TAGI also seamlessly incorporates additional structural attributes and does not require extensive preparation, making it better suited for modelling infrastructure deterioration based on visual inspections on a large scale.