Estimating corrosion growth rates for underground pipelines is a challenging problem. There are confounding variables with complex interaction effects that may result in unexpected outcomes. For instance, the relationship between soil conditions and AC interference is highly non-linear and challenging to model. This work expands upon prior work using a suite of machine learning tools to estimate corrosion rates. However, instead of estimating a single corrosion growth rate for a single girth weld address (GWA), this work estimates a distribution of potential corrosion growth rates. Modeling distributions provide a more effective risk-measurement framework, especially concerning high volatility or areas of severe tail risk. 

This work relies heavily on machine learning and geospatial tools - particularly artificial neural networks and gradient boosted trees to estimate the corrosion rates and non-linear processes. Building upon prior work using data from a North American Operator, the models in this paper use additional variables from recent research in AC interference and microbiologically influenced corrosion to construct a higher accuracy and distribution-based model of pipeline corrosion risk. 

Machine learning is a method that allows interpretations of new data using an established database of older data by referring to already-known results (known as “labels”) and extrapolating between them to estimate the label that would be assigned to a different experiment. This can be a powerful tool for corrosion prediction because it makes it possible to estimate a range of corrosion rates for a certain family of materials in a specific range of environments without actually performing experiments. In this paper the machine learning concept was applied to the erosion-corrosion of steels in white liquor a strongly alkaline industrial chemical used for pulping wood chips. Previously obtained corrosion data in white liquor which included different steel compositions particle concentrations and sizes temperatures and fluid properties such as viscosity were compiled and assigned labels based on previous assessments in the industry as passive acceptable marginal or unsuitable according to observed corrosion rate. Models using thirty selected variables were built based on this data using diverse machine learning methods including support vector machines (SVM) decision trees k-nearest neighbor methods (KNN). discriminant analysis etc. Feature selection was attempted for each model. The best accuracies for each method were compiled and assessed regarding their promise for predictive purposes in erosion-corrosion.

Visual inspection is a vital component of asset management that stands to benefit from automation. Using artificial intelligence to assist inspections can increase safety reduce access costs provide objective classification and integrate with digital asset management systems. The work presented herein investigates Deep Learning for automatic detection of corrosion on steel assets. A workflow is presented that spans from dataset creation to deployment highlighting the major hurdles and remaining work required. This process can be extended to other defects with a view to complete automation of visual inspection for asset management.