Pitting corrosion diagnostics and prognostics for miter gates using multiscale simulation and image inspection data

Physics-based high-fidelity pitting corrosion simulation models have successfully predicted the evolution of corrosion pit morphology for given mechanical and environmental conditions. However, applying such models for pitting corrosion diagnostics and prognostics in large civil infrastructures such as found in the inland waterways navigation enterprise is very challenging, primarily due to the impracticality of measuring individual pits. This paper addresses this challenge by bridging the gap between physics-based pitting corrosion simulation and vision-based pitting corrosion inspection of large civil infrastructures. The framework proposed in this paper consists of four main modules: mesoscale pitting corrosion simulation using the phase-field method, macroscale structural analysis, pitting corrosion detection using machine learning, and updating physics-based simulation models based on pitting corrosion detection. It begins with developing a forward simulation framework to predict the evolution of pitting corrosion on large civil infrastructure using multiscale analysis. A convolutional neural network (CNN)-based pit detection method is created in parallel to autonomously identify and extract pitting corrosion observations from corrosion inspection images. Finally, an approximate Bayesian computation numerical framework is proposed to update three key model parameters in the forward pitting corrosion simulation model using the detection results from the trained CNN model. The updated multiscale simulation model can then be used for pitting corrosion prognostics. A practical application example is demonstrated on miter gates to show the effectiveness of the proposed framework.