ABSTRACT

In a pipe, a circumferentially travelling ultrasonic wave will gather information about the properties and boundaries of the propagation medium. However, the compounded effects of diagnostic features like mean pipe wall thinning, surface roughness, regional depressions, and pit developments are difficult to separate using traditional methods. Therefore, this study proposes an approach using artificial neural networks to estimate the diagnostic features of interest.

This study is based on ultrasound simulations and synthetic data. The synthetic data is recorded at a set of transducer positions at the outer pipe wall. The resulting traces are then combined into 2D images where each vertical line represents the waveform recorded at a specific transducer location. The resulting images are used to train a neural network to extract relevant features.

Diagnostic features for mean and minimum thickness, as well as standard deviation of the wall thickness, are quite accurately estimated. The neural network-based estimation for mean thickness is more accurate than a conventional reference method. This is observed especially for non-uniform wall thickness, which is typically the case if the pipe wall has been exposed to erosion and corrosion. Features for depth and location of depressions are also informative but less accurate. Data decimation experiments have also been conducted, even down to one single remaining trace. Also in this case, the neural network is able to make good estimates of some features, especially the mean wall thickness.

INTRODUCTION

In a pipe, guided Lamb-like waves can propagate around the circumference of the pipe wall. As they do, the waves pick up details about the pipe wall's characteristics, such as its inner surface condition and, most significantly, its thickness. A robust pipe wall thickness estimation method based on conventional (i.e., non-machine learning) processing methods has been proposed by the authors [1]. However, the method fails to provide robust thickness measurements in cases with non-uniform wall thickness, which is typically the case when pipe suffers from inner corrosion and/or erosion. To improve the estimation of mean thickness in cases where the thickness is indeed non-uniform while also leveraging additional information embedded in the wavefield, this work turns to machine learning methods.

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