An Experimental Study of Predictive Metal Surface Corrosion via U-Net Segmentation and Random Forest Regression

Abstract

Corrosion of metal surfaces threatens the structural integrity and operational life of industrial assets, and there is a need to move from reactive inspection to proactive predictive maintenance. In this study, an intelligent monitoring framework is proposed for evaluating and predicting corrosion progression based on deep learning-based segmentation and ensemble-based regression. Pixel-level segmentation of corrosive features was performed using a U-Net convolutional neural network, which provided a quantitative basis for further analysis. The geographical data were then applied to Random Forest regression models to forecast future growth of the corrosion area and greyscale-based severity levels. According to the experimental results, the accuracy, precision and recall of the segmentation are 78.6%, 61.8% and 73.5%, respectively. The predicted models were found to be reliable with R2 values of 0.8477 for the growth in area and 0.9926 for increasing levels of severity. This study shows that computer vision along with predictive analytics provide a solid data-driven method for long-term asset management and threat prevention.