Prediction of atmospheric corrosion rate of steel using attention-augmented CNN and engineering application

Atmospheric corrosion is a critical driver of steel degradation in natural environments, and thus accurate assessment of the corrosion condition is essential for ensuring engineering safety. However, empirical models have limited applicability and poor transferability across different materials and environments. Traditional machine learning models, on the other hand, have weak expression and generalisation ability, along with poor interpretability. Therefore, an attention-augmented convolutional neural network (CNN) was developed. Firstly, a database of corrosion samples containing 24 types of steel, such as carbon, low alloy, and weathering steel, was established. Secondly, a one-dimensional CNN was formulated to extract complex high-dimensional features from sequential data, while two attention mechanism modules are incorporated to further enhance the representation of key information. Then, the interpretability of the model was analysed using the SHAP method. Finally, the long-term structural behaviour of a steel bridge was evaluated by coupling the intelligent model with numerical analysis framework. Results indicate the attention-augmented CNN model has higher accuracy than the existing models, with the evaluation metrics in test data being MAPE = 12.50%, RMSE = 5.45, MAE = 3.28, and R2 = 0.930, respectively. This study can provide a scientific basis for the long-term performance evaluation, and maintenance planning of infrastructure engineering.