Unsupervised deep neural network considering the uncertainties effect in pipeline condition monitoring using guided ultrasonic waves

Abstract

Pipeline condition monitoring is vital to ensure the safety of petrochemical pipeline systems. The Guided Ultrasonic Waves (GUW) method has been increasingly applied in pipeline condition monitoring as it provides reliable and accurate pipeline damage information. However, the existence of uncertainties from measurement and the finite element model may lead to unreliable or false damage identification. Attempts to deal with uncertainties, however, are only limited to large damage cases, while for small damage cases, the uncertainties may produce a greater amplitude than the reflected wave, resulting in unidentified reflected waves. Previous attempts employing unsupervised learning to compensate for uncertainties did not retain sufficient damage information for further damage localisation and quantification. Therefore, an unsupervised autoencoder deep neural network is proposed in this study to consider the effect of uncertainties by performing denoising while keeping the important features in the damage signal. The compensated and denoised output of the autoencoder is used to perform damage localisation and damage quantification using the proposed novel damage index based on the Residual Reliability Criterion (RRC). From the results obtained, the proposed method is able to perform uncertainty compensation and retain sufficient damage information for further damage localisation and damage quantification using the RRC damage index. It has shown consistent outcomes when the level of uncertainties is increased, even for low damage severity cases.