Bidirectional mapping modeling of pipeline vertical deformation and axial strain based on multi-source monitoring data and machine learning
Zhen Sun, Tianran Han, Xin Wang, Longxiang Wang, Hao Fu, Yalin Li, Ziqin Zhong, Jinlong Liu, Huang Huang, Zhishen Wu
Abstract
Axial strain and vertical deformation are critical indicators of pipeline structural integrity. Monitoring these parameters and constructing a bidirectional mapping model linking them enhance data completeness, system redundancy, and overall reliability. Thus, this study deployed embedded static leveling instruments and long-gauge fiber Bragg grating sensing systems in a pipeline project in Tangshan, Hebei Province. A total of 17,568 data sets were collected throughout 2024 to establish a comprehensive database. A GPKOA-XGBoost model was developed by integrating the Kepler optimization algorithm (KOA), good point set (GPS) strategy, and extreme gradient boosting (XGBoost) to achieve bidirectional mapping between strain and deformation. The sensitivity of model hyperparameters was assessed, and ablation experiments were performed to evaluate the contribution of each module. The influence of data size on model performance was also investigated. Furthermore, the performance of the GPKOA-XGBoost model was compared with 14 other machine learning models. Results indicate that GPS improves the population initialization of KOA, while KOA effectively tunes model hyperparameters. Their combined application enhances prediction accuracy. The model trained on six months of data achieved optimal performance by effectively balancing information content and noise. Among all models compared, GPKOA-XGBoost attained the highest R² and the lowest MAE, MAPE, RMSE, and MSE values. External validation with new data from 2025 yielded an R² exceeding 0.96, demonstrating excellent generalizability. A graphical user interface was developed based on this model to facilitate data recovery, cross-validation, visualization, and performance evaluation, thereby reducing technical barriers in engineering applications. In summary, the proposed approach provides an effective solution to improve data redundancy, data fusion efficiency, and reliability in pipeline monitoring systems.