The goal of this project is to leverage advances in machine learning and predictive analytics to advance the state of the art in pipeline infrastructure integrity management using forecasted (predictive) pipeline condition, using large sets of pipeline integrity data (periodic nondestructive inspection, NDI) and continuous operational data (e.g., sensor data used to monitor flow rate and temperature) generated by oil and gas (O&G) transmission pipeline operators.
Pacific Northwest National Laboratory (PNNL)
Natural gas and oil pipeline networks are critical infrastructures whose reliability is essential to sustaining energy sector operations and the U.S. economy. Establishing the reliability of transmission pipelines requires NDI after fabrication of the piping, and periodic inspection during operation. A range of existing and emerging sensor and instrumentation technologies are available for operational control, integrity assessment, and mitigation responses to off-normal events. At the same time, new sensors are being designed and proposed for many of the same measurement needs as well as for monitoring other parameters that could support and enhance operational decisions.
NDI techniques are typically used to detect the presence of degradation such as axial cracking, mechanical damage, or corrosion in oil and gas transmission pipelines. Typically, the detection of degradation triggers other analysis techniques, often based on structural mechanics, for assessing the structural integrity of pipelines. NDI techniques that have been applied for pipeline integrity assessment include magnetic flux leakage and ultrasound, both of which are used for in-line and offshore piping. Alternative UT (AUT) methods have been applied for inspecting girth welds in onshore and offshore piping. Alternative inspection methods include localized inspection techniques using visual, ultrasonic, guided wave ultrasound, eddy current, or enhanced visual inspection (such as dye penetrant inspection and magnetic particle inspection); however, these techniques often require physical access to the surface of the pipeline. Recent advancements have also resulted in 3D scanners for the rapid inspection of the external surface of the pipeline. Again, these require access to the surface of the pipe. Physical access may be a challenge if the pipe is buried or otherwise jacketed. Alternative emission techniques, visual observations, thermal imaging, chemical sensing, or process monitoring. Millimeter-wave and spectroscopic methods have also been proposed for detecting the presence of gaseous vapor due to a leak; however, these techniques are still in various stages of research and not commercially available.
The analytics to be developed will enable earlier detection of degradation in operating pipelines to support decisions about where to focus inspections, enhance situational awareness about the scope of inspections to be performed, allow monitor trending between inspections, and inform the timing of preventative maintenance.
Diagnostic Model: The team has been analyzing the first 6 batches of industry datasets (magnetic-flux leakage (MFL) signal data from a natural gas pipeline inspection service provider) and developing and enhancing a proof-of-concept CenterNet model, a fast and versatile image analysis model for object detection, and a Radial Basis Function model to predict the surface depth field of a defect which had been identified and centered by the CenterNet model.
The team has extracted the object detection component of CenterNet from the main source code and implemented the component for computational efficiency. The team used training data formats not natively supported by CenterNet and tested the model with an initial dataset of 80 flaws in a 24” pipe. Although the preliminary tests showed difficulty in identifying and classifying shallow flaws, the team continues to improve the detective precision by expanding the bounding box dataset to include the additional data at several pipe diameters.
The team also worked on developing a prototype Radial Basis Function model using an approach based on radial basis function representations to learn models which map MFL signals to the original defect geometry. Accurate predictions have been obtained of simplified flaws (e.g. hemi-spherical, cylindrical, rectangular) using these methods. So far, a simplified prototype model for the network was built and is currently being tested on the engineered defect dataset.
Prognostic Model: The team held a kick-off meeting with one natural gas pipeline operator on February 19, 2021 and a kick-off with a second operator on April 9, 2021. The data requirements document developed in 2019/2020 is being updated once more with specific details now that the NDAs are in place with the operators.
Federal Project Manager – Eric Smistad (firstname.lastname@example.org or 281-494-2619)
HQ Program Manager – Christopher Freitas (email@example.com or 202-586-1657)
Principal Investigator – Kayte Denslow (firstname.lastname@example.org or 509-375-2232)
Team Supervisor – John Duda (email@example.com)
Technology Manager – Jared Ciferno (firstname.lastname@example.org or 412-386-5862)