Event Details
October
26
2021
11
00
AM
to
October
26
2021
12
00
PM
(UTC-05:00) Eastern Time (US & Canada)
Microsoft Teams
Contact
Nathan Schoenkin
Nathan Schoenkin nathan.schoenkin@dot.gov

PHMSA: Low-Variance Deep Graph Learning for Predictive Pipeline Assessment with Interacting Threats

The Department of Transportation
Pipeline and Hazardous Materials Safety Administration
Pipeline Safety Research Program

 

Low-Variance Deep Graph Learning for Predictive Pipeline Assessment with Interacting Threats

Abstract Excerpt: This project was aimed to develop new low-variance data driven models and machine learning-based algorithms for characterization, simulation, diagnosis and prognosis of interacting threats in pipelines. Fundamental research was performed for proof of concept to utilize the advanced deep graph learning technology to assess the interacting threats which are more critical and could lead to premature failure of pipelines.


The hybrid and heterogeneous data-driven models were developed with a focus on simulating the interaction of corrosion anomalies and cracks. A machine learning-based spatiotemporal threat segmentation and matching algorithms were also developed for spatially segmenting threats and matching threats identified by sequential inspection data. A spatiotemporal deep graph learning model was developed to perform characterization and assessment of interacting threats, together with a Bayesian deep learning framework developed to estimate the uncertainty in the deep learning models.


The developed data-driven models can simulate magnetic flux leakage (MFL) and Pulsed Eddy Current (PEC) inspection data for interacting threat assessment. The developed spatiotemporal threat segmentation and matching algorithms can spatially segment threats from the images reconstructed from MFL and PEC data with the reduced variance to improve matching accuracy. The developed deep graph learning model was tested for diagnosis and prognosis of interacting threats. The models and methodologies developed in this project will provide the fundamental for future research and development on utilizing the advanced deep graph learning technology to improve the accuracy and reduce variances from currently used threat assessment methodologies for pipeline integrity management.


The project was performed by the joint research effort from Colorado School of Mines (CSM) and Michigan State University (MSU). The researchers from CSM completed the development of the threat segmentation and matching algorithms and the deep graph learning model, and the researchers from MSU completed the development of the data-driven models for simulation and characterization of interacting threats.


Project Link


Meeting Information
Tuesday October 26, 2021
11:00 AM ET


Microsoft Teams Meeting:
Click here to join the meeting


Agenda
Welcome and PHMSA Introductions
Project Presentation
Questions and Answers Period Adjourn