Monotonicity-constrained transfer learning for predicting residual tensile strength of GFRP under wind–sand erosion

Abstract

Accurate prediction of residual mechanical properties in fiber-reinforced polymer (FRP) composites under harsh environmental conditions is crucial for ensuring structural durability and service safety. This study proposes a monotonicity-constrained transfer learning ResNet (TL-ResNet) framework that integrates digital image correlation (DIC) strain field images with erosion parameters (erosion angle, erosion velocity, sand flow velocity, and exposure time) to predict the residual tensile strength of glass fiber-reinforced polymer (GFRP) laminates subjected to wind–sand erosion. This architecture fuses transfer learning, convolutional block attention modules (CBAM), and generalized mean (GeM) pooling to extract erosion-sensitive spatial features with precision under limited datasets. By embedding a monotonicity information regularization term in the loss function, it enforces monotonic degradation laws, ensuring monotonicity consistency in predictions and suppressing non-monotonicity oscillations. To enhance interpretability, the model integrates Grad-CAM, channel attention visualization, and SHAP analysis to reveal dominant erosion zones and parameter interactions influencing strength degradation. Experimental validation demonstrates performance metrics of R 2 = 0.8488, MAE = 45.30 MPa, and RMSE = 81.55 MPa. It outperforms traditional algorithms (KNN, RF, and XGBoost) and advanced models (CGCNN, CTGNN, and Multimodal Deep Learning) by 25%–40% in accuracy while reducing errors by 30%–50%. The proposed monotonicity-constrained framework successfully captures the experimentally observed monotonic degradation behavior of GFRP laminates under wind–sand erosion conditions. In addition, SHAP analysis provides interpretable insights into the relative contribution patterns of different input variables, indicating that erosion angle exhibits the highest influence on the prediction results, followed by erosion velocity, sand discharge rate, and erosion duration. This study establishes an interpretable monotonicity-constrained learning framework for predicting the long-term degradation behavior of composite materials under complex environmental erosion conditions.