Search

For safe operations and regular maintenance of pipelines in industrial and domestic usages, it is extremely important to extensively monitor gigantic pipeline systems and accurately detect harmful defects. Non-destructive evaluation (NDE) methods for inline inspection of these modern, massive pipe systems generate huge amount of data and need successful incorporation of automated defect detection algorithm that can conduct large-scale data analysis and provide instantaneous diagnosis by locating and identifying defects of varying characteristics. Here, we develop a novel NDE methodology for inspecting plastic pipelines used in natural gas distribution. Our proposed method is based on optical imaging and uses laser profiling to collect inline inspection data. In our laser profiling images, deformities due to defects lead to local perturbations in the circular rings produced by the otherwise non-defective inner walls of cylindrical pipes. Using experimental data collected by the proposed methodology, we studied the performance of a gamut of defect classification algorithms including conventional machine learning classifiers as well as deep learning architectures. Special attention was paid to take into account both the spatial and temporal features in our laser-scan image, particularly exploiting the circular shape of non-deformed laser rings. This led to the development of a hierarchical bilinear pooling (HBP) based deep learning framework that can precisely identify local perturbations in our laser-scan data by assimilating information across multiple cross-layer features. Equipping our endoscopic laser profiling system with the HBP based defect classification algorithm, we develop an integrated diagnosis method for inline data collection and automated defect characterization. Based on experimental data collected under varying profiling conditions, we demonstrate superior performance of our proposed integrated method in damage localization and classification. [ABSTRACT FROM AUTHOR]