A two-stage pipeline defect depth estimation method based on MFL signals

Magnetic flux leakage (MFL) internal inspection is widely used in pipeline non-destructive testing (NDT) with its low cost and high efficiency. Affected by different environments, MFL signals are complex and changeable, which has brought severe challenges to the accurate estimation of defect depth. In this paper, a novel pipeline defect depth estimation method is proposed to improve the accuracy, which can extract deep features from MFL signals adaptively. First, an auto-encoder is designed to extract primary features from MFL signals in an unsupervised way. Second, a classification neural network is developed to extract secondary features in a supervised way. To make better use of the feature information, part of the weight and structure from auto-encoder is reused in this network. Third, a random forest model is constructed to study the complex nonlinear relationship between features and defect depth. Experimental results are compared with state-of-art method from seven aspects, which highlight the superiority of the proposed method in defect depth estimation.