Machine Learning-Based Severity Assessment of Pipeline Dents

One challenge to pipeline operators is to identify potentially injurious dents among thousands of reported deformation features using limited information (e.g., reported dent’s length, width, and depth) and to prioritize the efforts and allocate the resources to obtain additional more detailed information (e.g., dent profiles) for those potentially severe dents. An innovative approach based on machine learning predictions stemming from a representative dictionary of finite element analysis (FEA) generated prototypes was developed. The proposed approach predicts multiple severity-based indicators for each dent, then combines them in an overall severity score, which finally is used to prioritize the acquisition of dent profiles. Once the dent profiles are available, detailed level 3 FEA quantitative reliability analyses, following previously developed and published methodology (QuAD) [1], is performed allowing pipeline operators to confirm dent’s severity more accurately and perform an integrity risk informed decision (IRIDM) leading to a safer and more efficient integrity management. Three severity indicators were considered herein and intended to address both formation-induced and service-induced failure mechanisms. The maximum dent formation plastic strain and accumulated ductile failure damage were used for evaluating the likelihood of forming a crack during indentation. The third indicator was the stress concentration factors (SCFs) to assess the potential of service-induced failure due to fatigue. A machine learning model, as an emulator, trained and tested using ∼4000 FEA-based dent prototypes was shown to be able to effectively predict dent severity indicators previously referred to. These predicted dent severity indicators are combined to produce an overall severity score, which was finally used to prioritize the acquisition of the detailed dent profiles. Once profiles are obtained, detailed FEA quantitative reliability assessments will ultimately confirm the severity and hence drive repair/no repair decisions, enabling in this way an efficient and effective allocation of resources.