Fusing physics-inferred information from stochastic model with machine learning approaches for degradation prediction

Many methods have been developed for degradation prediction. Machine learning-based methods have the advantages of capturing complex non-linear relations in degradation processes, but they often suffer from limitations of available dataset. Physics-based methods are closely related to physical degradation mechanism, while they are often limited by the incompleteness of modeling. This work presents a new hybrid method for degradation prediction with bias correction by combining the degradation tendency information from the physics-based stochastic degradation model with machine learning approaches. A gamma-gamma two-stage degradation model is adopted to obtain the expected degradation path, which is used to develop the input of Bi-direction long-short term memory (Bi-LSTM) network for degradation prediction. Case studies of bridge deck rebar degradation are conducted to demonstrate the applicability of the proposed approach, where actual data is collected based on 33-months of rebar degradation experiments under different environmental conditions. The results show that the proposed hybrid method outperforms other machine learning-based methods. Specifically, the mean square error of rebar degradation prediction is reduced at least by 27% under the proposed approach in comparison with pure Bi-LSTM. This work provides insights for performing bias learning in prognosis by leveraging the advantages of physics-based methods and machine learning approaches.