This study focuses on applying machine learning algorithms to predict the corrosion depth of facility station piping assets, as well as comparing the computational accuracy of the predicted corrosion depth based on various machine learning algorithms. Simulated corrosion testing data of facility piping was fit into the following machine learning algorithms: Gradient Boosting(GBM), Artificial Neural Network (ANN), and Random Forest (RF). K-fold cross validation was used to evaluate the models and grid search was applied for the models to refine and calibrate each model. The variable sensitivity analysis was conducted separately for the external and internal corrosion of station piping, and it assisted in limiting the number of independent variables included in machine learning models. This study compares the performance of corrosion depth prediction models for facility station piping and draws conclusions on model performance based on performance evaluation metrics.


The pipeline industry has widely used integrity principles to manage time-dependent and time-independent threats. The detection of time-dependent threats such as corrosion has been accomplished by using inline inspection tool technologies such as ultrasonic and magnetic flux leak inspection tools. However, most facility piping assets can not easily be inspected using in-line inspection methods and must instead be assessed using data collected from operations, such as flow frequency, product type, Cathodic protection record, or Direct Assessment Methods using Non Destructive Testing such as ultrasonic measurements or monitoring of corrosion coupons. This only provides limited data at discrete locations and in the case of ultrasonic measurement can be very costly due to coating removal and/or excavation. Corrosion is highly complex due to a variety of factors including, types of products being transported, flow conditions that can remove protective scaling layers and accelerated galvanic activity, conditions of the surrounding ground, and microbiological mechanisms. While there is significant work in the area of mechanistic models that can estimate corrosion feature initiation and growth, there still can be considerable variability in observed corrosion growth that is not captured by any one model, which makes the implementation of mechanistic models challenging. There is a need for numerical inspection tools to assist in understanding the severity of corrosion on the pipe to prioritize the risk and hazard location.

This content is only available via PDF.
You can access this article if you purchase or spend a download.