Pitting Detection and Characterization From Ultrasound Timelapse Images Using Convolutional Neural Networks

Abstract

Pitting corrosion, a localized form of corrosion leading to cavities and structural failure in metallic materials, requires early detection for effective mitigation. While ultrasonic inspection techniques can readily detect uniform wall thinning, they often struggle to identify pitting corrosion. This study proposes a time-lapse ultrasound inspection method to detect early-stage pitting using pulse-echo sensors. By recording multiple ultrasonic traces over time, 2-D timelapse images of ultrasonic reflectivity can be generated and fed into a trained neural network for pitting diagnostics. In general, training a machine-learning model requires a large training dataset. This work used data from a drilling experiment to generate a suitable dataset. Dataset construction by random time-ordered combinations of ultrasonic measurements was conducted to create a diverse set of time-lapse image samples to generalize the resulting machine-learning model adequately. A classification neural network was trained to detect the presence of drilled holes, and a separate regression network was trained to estimate the hole depth. Based on drilling data from an independently acquired test dataset, results demonstrate a mean absolute error of 0.163 mm for hole depth estimations. All holes are successfully detected when 0.1 mm deeper than the defined pitting threshold of 0.5 mm. This suggests that the proposed method generalizes well and can be deployed to any similar acquisition system.