Residual strength prediction of corroded pipelines based on physics-informed machine learning and domain generalization
Tingting Wu, Xingyuan Miao, Fulin Song
Abstract
Accurate residual strength prediction of corroded pipelines is of great significance to pipeline reliability assessment and transportation safety. In this paper, a novel physics-informed domain generalization model is proposed for predicting the residual strength of corroded pipelines. Firstly, a physics-informed feature space is constructed with physical prior knowledge. Secondly, Gini coefficient is utilized to rank the feature importance. To determine the optimal feature subset, the feature number, prediction accuracy and model stability are comprehensively considered. Subsequently, a predictive model based on deep forest (DF) and reinforcement learning is developed, double deep Q-network (DDQN) is used to optimize the hyper-parameters. Finally, the domain generalization performance is verified for different application scenarios. The results demonstrate that the proposed model has better prediction performance than other models, and generalization performance is the best in term of external loads and multiple defects. This study provides a novel perspective for integrity assessment of corroded pipelines.