Automated Pipeline Diagnostics Using Image Processing and Intelligent System

Due to many drawbacks such as human error, tremendous energy and time consumption in traditional method of pipeline inspection, this paper proposes an automated pipeline diagnostic using image processing and intelligent system. The primary focus of the developed system is underwater pipeline network due to higher inaccessibility and defect rate. Comparatively, many methods were used in image processing along the years and Convolutional Neural Network (CNN) was identified as the most effective method for this case study based on the literature review. Narrowing down into CNN context, the author has identified and compared the mean accuracy of transfer learning process of two pre-trained convolutional neural networks which were also the winners of ImageNet Large Scale Visual Recognition Challenge (ILSVRC) for the year 2012 and 2014. They are known as AlexNet and GoogLeNet. This was done by initially modelling pipes with various defects in CATIA and surface recording was simulated similar to ROV recording. Then these videos were automatically converted into image frames, pre-processed and fed into system as training material. After sufficient training, the system was able to detect and distinguish the pipeline defects. GoogLeNet was identified as the network with the highest mean accuracy of 99.87, hence was finalised as the systems network architecture. MATLAB 2017b was used to develop the system. To further evaluate the performance of the system, a mini lab rig was set up replicating underwater environment with pipeline models with dents, holes and cracks. Similarly, the inspection videos were recorded and the system was able to detect and distinguish the defects on the pipeline alongside their location and percentage coverage with mean accuracy of 99.87 as well, proving the functionality of the system in real condition. The mechanical properties of the pipelines and characterisation of pipeline defects were also reviewed thoroughly before developing the inspection system. © 2020, Springer Nature Singapore Pte Ltd.