An adaptive hybrid deep learning-based reliability assessment framework for damping track system considering multi-random variables

Abstract

Reliability assessment is essential in the design and management of rail transit system (RTS), with a heightened focus on damping tracks that feature low-stiffness structures and various uncertainties. However, existing research on RTS has notable gaps: one is the unexplored nonlinear track irregularities in surrogate model, while the other is the significant computational burden of current reliability analysis. Therefore, Firstly, based on a database acquired from simulations of rigid-flexible coupling dynamic models, a hybrid deep learning (DL) surrogate model is developed and trained. Notably, it accommodates both time-series and feature parameters data as inputs, and integrates a novel algorithm featuring slide window with variational decomposition mode-sample entropy, optimization algorithm and adaptive learning strategy (AS). Subsequently, an advanced reliability assessment framework is proposed, utilizing AS hybrid DL-based Probability Density Evolution Method (PDEM), which thoughtfully designed to analyze the two aforementioned categories of random variables and then applied to a train-steel spring floating slab track. Experimental studies show the proposed surrogate model effectively and robustly predicts four safety metrics, with each component's excellence confirmed. Furthermore, this framework shows higher accuracy than the Monte Carlo Method and enhances computational efficiency by 10 ∼ 30 times compared to traditional mechanism-based PDEM. Consequently, reliability assessments indicate values of 0.9312 and 0.9214 for wheel-load reduction rate and rail displacement, respectively, with other metrics showing zero safety hazards. Findings of this study have practical applicability in the field of RTS design and maintenance.