Near-And Remote-Field Eddy Current Data Fusion: Wellbore Casing Inspection with Hybrid Neural Networks

Abstract

Abstract The structural stability of wellbores depends on the concentric steel casings that are lowered into the wells and cemented in place. Such casings are often subjected to intense forces and high pressure, as well as being exposed to corrosive elements. As a result, defects such as pits, cracks, and other forms of metal loss inevitably occur on the casings. The presence of defects poses a threat to wellbore integrity that increases overtime as the metal losses increase in both depth of penetration and surface area, which may result in severe environmental and financial damage if left unchecked. Hence, many acoustic, visual, and electromagnetic (EM) inspection methods have been developed to assess the health of casings to facilitate risk management decisions. EM inspection methods are widely used because of their ability to detect metal loss on multiple concentric casings while being largely unaffected by the cement between the casings. While visual and acoustic methods generally produce results that are readily interpretable, EM measurements are often more difficult to utilize due to their high nonlinearity. This research investigates the EM inspection of wellbore casings using the near- and remote-field eddy current (NFEC and RFEC) methods. Cross-sectional images are reconstructed by a hybrid neural network (HNN) with two parallel modules that map EM measurements to the pixels of the images. A specialized neural network module is designed for each of these methods. Both modules include convolutional and recurrent layers in their structures to extract spatial and sequential attributes from EM data. Using this approach, the physical locations of metal loss and casing material are inherently represented by the coordinates of the pixels on the reconstructed image, while the values of the pixels represent the probability of metal loss at their location. In addition, in-depth analyses show that this approach is generalizable to metal loss scenarios that are different in terms of shape and location from the training data.