2024 – 20

Emmanuel Valencia, Atul Ganpatye

ADV Integrity, Inc., Magnolia, USA

Abstract

Traditional dent assessments involve calculating deformation strains and stress concentration factors (SCF) as proxies for the severity of the dent. Typically, the determination of SCF requires advanced computational methods (like finite element analysis) with specialized software and skillsets. This can be time-consuming and subjective, greatly depending on the engineer's skill level using the approach. Of the two factors described above, deformation strain calculations have been well-established and definitive because strains can be directly and readily interpreted from the deformation geometry. Therefore, this paper only focuses on the SCF aspect where the proposed approach, using machine learning (ML), can significantly reduce the efforts and resources needed to determine the SCF. Due to the non-linear relationship of the dent shapes and their corresponding SCF values, multiple deep learning ML methods were considered in this study. The final model architecture involved Convolutional Neural Networks (CNN) due to their ability to extract features, group objects, and discover valuable data patterns from unknown elements. The CNN model was trained on 4,667 raw ILI data files containing dent shapes which underwent significant preprocessing, such as data normalization, filtering, and smoothing. The final model was selected after careful fine-tuning of the hyperparameters and comparison of 180 CNN model variations where it successfully predicted SCFs ranging from 1.04 to 10.69 with a root mean squared error (RMSE) of 0.418 and a coefficient of determination (R2) of 0.929, indicating a high level of accuracy and goodness of fit, respectively. Although the results are favorable, the training dataset is biased toward a single pipeline operator and therefore does not represent the global statistics of dented pipelines. Along these lines, the paper discusses ideas and proposals for a more robust implementation of the approach for wider applications. The work demonstrates that the model has the potential to serve as an assessment tool to provide immediate, in-the-ditch guidance to pipeline operators, therefore reducing downtime and limiting resources.

Read/Download PaperGrade