Denoising autoencoder in damage detection of pipeline using guided ultrasonic wave

Abstract

Pipeline condition monitoring is essential in critical sectors such as the petrochemical, nuclear and energy sectors. The guided ultrasonic wave (GUW) monitoring system is an available pipeline condition monitoring system that is gaining much attention owing to its portability, long coverage and high sensitivity to damage. However, environmental and operational conditions (EOCs) effects, especially temperature and random noise may generate unwanted peaks, which are falsely identified as damage. Attempts to deal with EOC effects have not solved the problem, especially for small damage cases (damage equal to or less than 5% cross sectional area loss (CSAL)). In this study, a new damage feature extraction method based on the residual reliability criterion (RRC) is proposed. The performance of the proposed method is measured using the established receiver operating characteristics (ROCs) performance evaluation method. The findings show that this method performs well, with an AUC value greater than 0.9, based on numerical model under 40 ? variations and 10% random noise level, and that the application of RRC is intuitively simple. To ensure the practicality of the method, a 6 metre long, 8 inches diameter experimental pipe model filled with liquid is used to form a GUW database of small damage under 30 ? variations by using Torsional T(0,1) excitation mode at 26 kHz centre frequency. However, the RRC underperformed when experimental data is used because the random noise generated by healthy and damaged signals interferes and generates high amplitude noise. Therefore, this study proposed a denoising autoencoder (DAE) neural network to deal with the effects of EOCs. A DAE decodes high-dimensional data into low-dimensional features and reconstructs the original data from these low-dimensional features. By providing GUW signals at a reference temperature, this structure forces the DAE to learn the essential features hidden within complex data. The proposed DAE showed perfect detection (AUC value of 1.0) using numerical model and performs well (AUC greater than 0.9) using experimental model in terms of small damage identification. Moreover, the proposed method showed superiority among other advanced EOC compensation techniques using both numerical and experimental models.