Experimental investigation and data-driven ANN modeling of failure pressure in GFRP-repaired pipelines

The failure pressure of glass fiber-reinforced polymer (GFRP)-repaired steel pipelines is investigated using an integrated experimental result with a developed artificial neural network (ANN) model for predicting the output. The experimental part involved pressurized blister tests. It was performed on mild steel plates, containing circular through-wall defects of varying diameters (10–25 mm) and repaired with GFRP laminates of different thicknesses (8–32 mm). To investigate the failure pressures, there are five mains defects’ parameters such as diameter, repair thickness, fiber orientation, repair size, and roughness, with a total of 60 experimental runs to establish a comprehensive dataset. The ANN model employed a feedforward structure trained using the backpropagation learning algorithm with two hidden layers and optimized through hyperparameter tuning. The N-2 model is best network over five developed models (N-1 to N-5), characterized by a logsig–tansig activation combination, Bayesian Regularization (trainbr) optimization function, and six neurons. The best predictive performance was achieved with R² = 0.969 and MSE = 1.06 MPa². Thus, the findings showed high alignment between experimental and predicted values to confirm the model's ability to acquire nonlinear relationships among repair parameters. Additionally, the defect diameter and repair thickness are the significant factors affecting failure pressure using the sensitivity analysis, while the other factors have less effect on the output. The model implementation attributes an accurate and scalable predictive approach for evaluating FRP-repaired pipelines in the oil industry. This achievement contributes to a foundation for future digitalization systems for predicting the maintenance strategies in pipeline integrity management at a low cost.