Two-stage Degradation Modeling Combined With Machine Learning For Steel Rebar Degradation Prediction

• Main Authors: Zhou, Jian, Coit, David. W., Nassif, Hani, Li, Zhanhang
• Format: Conference Proceeding
• Language: English
• Subjects: Bridges; Degradation; Machine learning; Machine learning algorithms; Prediction algorithms; Predictive models; reliability; steel rebar corrosion prediction; Stochastic processes; two-stage degradation model
• ISSN: 2577-0993
• DOI: 10.1109/RAMS51457.2022.9894005

Degradation models are becoming important for analytically investigating degradation behaviors of industrial systems and products. Many real-world degradation processes show distinct stages. A two-stage stochastic degradation model with degradation initiation is proposed. The model contains a time-to-event distribution describing degradation initiation latency in the first stage, and a stochastic process depicting degradation propagation in the second stage. A hybrid framework using the degradation tendency obtained from the two-stage degradation model to construct the input and output for the machine learning (ML) algorithms for degradation prediction is also developed. Based on experimental data, the impacts of pre-cracking condition and chloride solution concentration, as important field environment stresses, on the degradation processes of different steel rebars embedded in reinforced concrete bridge decks are investigated. In the case studies, Weibull distribution is adopted for the first degradation stage and an adjusted gamma process is developed for the second degradation stage. The two-stage degradation model has an advantage to help design accelerated degradation tests considering the influences of environmental stresses on each stage. It remarkably improves the efficiency of estimating bridge rebar reliability and its expected degradation. The model also helps analyze the stress-degradation relationship. The hybrid framework outperforms pure ML methods on degradation prediction proving the performance gains by leveraging the advantages of physics-based degradation model and ML algorithms.