Development and fusion of NDT classifiers for defect detection on underwater structures

Abstract

Vessel hull inspection is a complex and labor-intensive task, motivating the utilization of data fusion from complementary nondestructive testing (NDT) techniques to boost reliability and reduce detection uncertainties. In this study, an underwater robot equipped with three NDT systems – monocular RGB camera, eddy current (ECT), and ultrasound (UT) – is employed to assess a 12.7 mm thick ASTM-A36 carbon steel plate with artificially induced defects – coating loss, crack, pitting, and generalized corrosion. A classifier is developed for each NDT system for defect classification. Mask R-CNN is used for the camera data, achieving a mean average precision of 67.7% across three defect classes. A multilayer perceptron is used to classify the ECT data into six defect types and the non-defective class, attaining an accuracy of 98.8% on the test set. For UT data, a rule-based classifier based on A-scan-derived features distinguishes four defect classes. Subsequent to spatial synchronization, these classifier outputs are fused at the decision level, where redundancies between the camera and ECT classifiers are eliminated. Their outcomes are concatenated with the UT classifier results, culminating in a comprehensive detection map capable of representing eleven distinct defect classes. The classifiers and fusion are evaluated on a novel plate with unseen artificial defects. The camera and UT classifiers excel in detecting all classes for which they were specifically trained. In contrast, ECT exhibits a more limited capability, detecting only 21% of the specified defect classes. The findings underscore the potential of NDT data fusion to enhance inspection reliability by minimizing redundancies and extending the detection range.