Assessment model for the long-term behavior of PSC-I girder using a Bayesian physics-informed neural network

초록

The residual prestress in PSC girders is a critical factor governing the serviceability and strength of the member; however, its in-service evaluation remains challenging, and continuous long-term monitoring of many girders through sensing-based approaches is impractical under typical field conditions. This paper proposes a Bayesian physics-informed neural network (B-PINN) framework that estimates the long-term prestress loss of PSC-I girders from their deformation data. The age-adjusted effective modulus method (AEMM) with the relaxation approach is embedded into the PINN as the physics base through the curvature-displacement differential equation. An inverse problem integrates sparse camber observations-augmented by logarithmic regression-into the network, and particle swarm optimization recalibrates the creep coefficient against the re-estimated curvature. The framework is extended to a Bayesian setting via Monte Carlo simulation of the concrete elastic modulus and initial prestress, combined with Hamiltonian Monte Carlo sampling, yielding posterior distributions of girder curvature, deflection, and prestress loss. Validated through two case studies on experimentally monitored posttensioned concrete beams, the B-PINN posterior consistently brought the predicted long-term stress loss closer to the measured response than the AEMM baseline, demonstrating its validity as a probabilistic assessment scheme for the residual prestress of in-service PSC-I girders based on measured displacement data.

키워드