Spatial growth prediction for filiform corrosion: classification models with image-derived features

The Government Accounting Office reports that the Department of Defense spends approximately $20 Billion per year on prevention and repair of corrosion. The resources required to combat corrosion problems are desperately needed for equipment and personnel in today’s high operations tempo environment. Despite its significance, modeling for the prediction of corrosion evolution has been largely ignored in the literature. The purpose of this research is to determine iffeatures derived from images of corrosion growth can provide insight into the corrosion process. Feature data was derived from time sequenced images of filiform growth in Al2024-T3. Logistic regression, classification trees, generalized additive, and kernel density models were considered for predicting corrosion growth from day 1 to day 2 of exposure. A model for random growth was developed to provide a metric for evaluating the model performance. The logistic regression, classification tree, and generalized additive models all had similar performance in predicting corrosion growth for the training data. The models had similar performance on the test data with approximately 70% recall and precision. However, when compared to the random model, they offer no improvement in predictive performance. The features derived from the images of corrosion growth are insufficient to describe the corrosion growth process. Although these images are readily available to maintainers performing inspection of corrosion damage, smaller scale features should be considered to provide a more accurate prediction.