1RSI Pipeline Solutions, New Albany, USA. , 2NASA, Cleveland, USA. , 3Pacific Gas and Electric, Oakland, USA
1RSI Pipeline Solutions, New Albany, USA. 2NASA, Cleveland, USA. 3Pacific Gas and Electric, Oakland, USA
ABSTRACT
Pipeline operators in the United States are increasingly relying upon materials verification programs (MVP) to establish the properties of pipelines lacking reliable records. The ongoing MVP at the Pacific Gas and Electric Company (PG&E) applies external nondestructive testing (NDT) to exposed line pipe to gain insight into the grade, vintage, and manufacturing method of the pipe. PG&E supplements the standard NDT methods for composition, strength, and geometry with the nondestructive collection of surface microstructures using metallographic replicas. The microstructures are quantitatively evaluated for ferrite grain size and fraction of pearlite. These are then used in conjunction with other measured characteristics to support determination of grade and vintage, and to identify populations of similar pipes. Automating these analyses is of interest because the manual evaluations are labor intensive and subject to variability associated with evaluator skill, judgement, and fatigue.
Traditional methods for automating microstructure analyses are often challenged by small variations in sample or image quality. Machine learning (ML) models have been shown to be more robust, but training these models typically requires hundreds of manually pre-processed images. This creates a high initial investment that impedes practical implementation in an operational environment. Recently, pre-training ML models on a large number of generic images has been shown to substantially reduce the required number of application-specific training images.
This work will describe the performance of an open-source ML model pre-trained on a database of over 105 microscopy images and subsequently trained on fewer than 20 line pipe microstructures. The training and validation of the model will be presented, along with a comparison of ML and manual evaluations of more than 150 microstructures from more than 50 line pipes. The results will show the performance of the ML model to be comparable to that of manual evaluations.



