Structural health monitoring on pipeline system using unsupervised machine learning algorithm

Abstract

Pipeline network system has been the most vital infrastructure for several needs ranging from residential, industrial, oil and gas, aerospace, automotive and many more. However, such system is also vulnerable to defects at some point during its lifespan. Therefore, a proper structural health monitoring on the pipeline system is vital to ensure optimum safety of users and efficient transportation of liquid and gas. The objectives of this study are to investigate the natural frequency of the pipeline, establish a machine learning algorithm for pipeline damage detection and demonstrate numerically the applicability of machine learning in defect identification on a pipeline. A total of 4 single long pipes pinned at both ends are modelled using ABAQUS software whereas the K-mean algorithm is built via Google Colaboratory. The pipes are in the form of 2D wire with no loads applied. The pipes are categorized into 2 natures i.e healthy and corroded and are partitioned into 4 parts. The corrosion is induced on 3 out of the 4 pipes specifically at one of the portioned parts prior to undergoing frequency analysis to acquire mode shapes with their respective natural frequencies. As for the algorithm, 2 clusters, 0 and 1 are determined and labelled as healthy and corroded respectively. Multiple mode number ranging from 0 to 11 that represent the range of 4 distinct natural frequency data for 4 different pipes are fed into the algorithm and classified based on the pre-determined clusters. Based on the results obtained, the presence of corrosion on the pipe influences the deformation of the pipe by imposing slightly higher natural frequency in the range of 1.03% to 10.4% and 2 out of 4 pipes with damage locations at 1 and 3 provide identical natural frequency. The algorithm exhibits inaccurate damage detection as it manages to identify two damage locations at 1 and 3 when only one mode number is fed but eventually provides accurate damage detection for all 3 locations when more than one mode number. However, due to the identical natural frequency for location 1 and 3, the damage localization cannot be performed by the algorithm. As a conclusion, the competency of K-mean clustering in defect identification has achieved a satisfactory remark with the exception of damage localization.