Efficient Training Architecture for Circumferential Weld State Prediction Model and Hyperparameter Co-evolution Strategy

To enhance the performance and reliability of engineering decisions in the quality prediction of in-service natural gas pipeline girth welds, this study conducted in-depth optimizations on existing ensemble learning models. Based on a structured database constructed from multi-source data—including welding process parameters during construction, as well as corrosion detection, internal inspection (radiographic/magnetic particle/phased array ultrasonic), and excavation verification data collected during operation and maintenance—the core optimization efforts focused on fine-tuning the hyperparameters of the CatBoost and XGBoost models using the Optuna framework. This process, known as hyperparameter co-evolution, involves the simultaneous optimization of multiple hyperparameters to improve model performance. Among these, the CatBoost model demonstrated superior performance. Furthermore, by integrating the SHAP machine learning interpretation framework with the point selection criteria outlined in the "Code for Excavation Verification of Internal Inspection of Steel Pipelines," the interpretability of the model's decision-making mechanism was significantly improved. This approach quantitatively revealed the significant influence of key features such as welding process parameters, service duration, and radiographic inspection defect depth. The optimized model was rigorously validated through engineering applications, demonstrating that its key performance indicators—particularly accuracy in excavation inspection point selection—substantially outperformed traditional empirical methods. Through the synergistic enhancement of model performance and interpretability, this study offers robust, data-driven decision support for managing quality risks associated with in-service pipeline girth welds.